Release 0.24.0

  -- A long awaited rework with floating point images, modern Rust,
  simplified traits, and more image formats.

Breaking changes

Structural changes:
- Minimum Rust version is now `1.56` and may change in minor versions until
  further notice. It is now tracked in the library's `Cargo.toml`, instead, by
  the standard `[package.rust-version]` field. Note: this applies _to the
  library itself_. You may need different version resolutions for dependencies
  when using a non-stable version of Rust.
- The `math::utils::{nq, utils}` modules have been removed. These are better
  served through the `color_quant` crate and the standard library respectively.
- All codecs are now available through `image::codecs`, no longer top-level.
- `ExtendedColorType` and `DynamicImage` have been made `#[non_exhaustive]`,
  providing more methods instead of exhaustive matching.
- Reading images through the generic `io::Reader`, as well as generic
  convenience interfaces, now requires the underlying reader to be `BufRead +
  Seek`. This allows more efficient support more formats. Similarly, writing
  now requires writers to be `Write + Seek`.
- The `Bgra*` variants of buffers, which were only half-supported, have been
  removed. The owning buffer types `ImageBuffer` and `DynamicImage`
  fundamentally already make a choice in supported pixel representations. This
  allows for more consistent internal behavior. Callers are expected to convert
  formats when using those buffers, which they are required to do in any case
  already, and which is routinely performed by decoders.

Trait reworks:
- The `Pixel` trait is no longer implemented quite as liberally for structs
  defined in the crate. Instead, it is now restricted to a set of known channel
  which ensures accuracy in computations involving those channels.
- The `ImageDecoderExt` trait has been renamed to `ImageDecoderRect`, according
  to its actual functionality.
- The `Pixel` trait and its `Subpixel` field no longer require (or provide) a
  `'static` lifetime bound.
- The `Pixel` trait no longer requires specifying an associated, constant
  `ColorType`. This was of little relevance to computation but made it much
  harder to implement and extend correctly. Instead, the _private_
  `PixelWithColorType` extension is added for interfaces that require a
  properly known variant.
- Reworked how `SubImage` interacts with the `GenericImage` trait. It is now a
  default implementation. Note that `SubImage` now has _inherent_ methods that
  avoid double-indirection, the trait's method will no longer avoid this.
- The `Primitive` trait now requires implementations to provide a minimum and
  maximum logical bound for the purpose of converting to other primitive
  representations.

Additions

Image formats:
- Reading lossless WebP is now supported.
- The OpenEXR format is now supported.
- The `jpeg` decoder has been upgraded to Lossless JPEG.
- The `AvifEncoder` now correctly handles alpha-less images. Some additional
  color formats are converted to RGBA as well.
- The `Bmp` codec now decodes more valid images. It can decode a raw image
  without performing the palette mapping. It provides a method to access the
  palette. The encoder provides the inverse capabilities.
- `Tiff` is now an output format.

Buffers and Operations:
- The channel / primitive type `f32` is now supported. Currently only the
  OpenEXR codec makes full use of it but this is expected to change.
- `ImageBuffer::{get_pixel_checked, get_pixel_mut_checked}` provide panic-free
  access to pixels and channels by returning `Option<&P>` and `Option<&mut P>`.
- `ImageBuffer::write_to` has been added, encoding the buffer to a writer. This
  method already existed on `DynamicImage`.
- `DynamicImage` now implements `From<_>` for all supported buffer types.
- `DynamicImage` now implements `Default`, an empty `Rgba8` image.
- `imageops::overlay` now takes coordinates as `i64`.

Limits:
- Added `Limits` and `LimitSupport`, utilized in `io::Reader`. These can be
  configured for rudimentary protection against resource exhaustion (images
  pretending to require a very large buffer). These types are not yet
  exhaustive by design, and more and stricter limits may be added in the
  future.
- Encoders that do provide inherent support for limits, or reserve a
  significant amount of internal memory, are urged to implement the
  `set_limits` extension to `ImageDecoder`. Some strict limit are opt-in, which
  may cause decoding to fail if not supported.

Miscellaneous:
- `PNMSubtype` has been renamed to `PnmSubtype`, by Rust's naming scheme.
- Several incorrectly capitalized `PNM*` aliases have been removed.
- Several `enum` types that had previously used a hidden variant now use the
  official `#[non_exhaustive]` attribute instead.

Release 0.23.14

- Unified gif blending in different decode methods, fixing out-of-bounds checks
  in a number of weirdly positioned frames.
- Hardened TGA decoder against a number of malicious inputs.
- Fix forward incompatible usage of the panic macro.
- Fix load_rect for gif reaching `unreachable!()` code.

- Added `ExtendedColorType::A8`.
- Allow TGA to load alpha-only images.
- Optimized load_rect to avoid unnecessary seeks.

sha256sum:
24ffcb7e7244a9bf19d35bf2883b9c080c4ced3c07a9895572178cdb8f13f6a1  target/package/image-0.23.14.crate

Release 0.23.13

- Fix an inconsistency in supported formats of different methods for encoding
  an image.
- Fix `thumbnail` choosing an empty image. It now always prefer non-empty image
  dimensions.
- Fix integer overflow in calculating requires bytes for decoded image buffers
  for farbfeld, hdr, and pnm decoders. These will now error early.
- Fix a panic decoding certain `jpeg` image without frames or meta data.
- Optimized the `jpeg` encoder.
- Optimized `GenericImage::copy_from` default impl in various cases.

- Add `avif` decoders. You must enable it explicitly and it is not covered by
  our usual MSRV policy of Rust 1.34. Instead, only latest stable is supported.
- Add `ImageFormat::{can_read, can_write}`
- Add `Frame::buffer_mut`
- Add speed and quality options on `avif` encoder.
- Add speed parameter to `gif` encoder.
- Expose control over sequence repeat to the `gif` encoder.
- Add `{contrast,brighten,huerotate}_in_place` functions in imageproc.

- Relax `Default` impl of `ImageBuffer`, removing the bound on the color type.
- Derive Debug, Hash, PartialEq, Eq for DynamicImage

sha256sum:
293f07a1875fa7e9c5897b51aa68b2d8ed8271b87e1a44cb64b9c3d98aabbc0d  target/package/image-0.23.13.crate

Version 0.23.12

- Fix a soundness issue affecting the impls of `Pixel::from_slice_mut`. This
  would previously reborrow the mutable input reference as a shared one but
  then proceed to construct the mutable result reference from it. While UB
  according to Rust's memory model, we're fairly certain that no miscompilation
  can happen with the LLVM codegen in practice.
  See 5cbe1e6 for more details.
- Fix `imageops::blur` panicking when `sigma = 0.0`. It now defaults to `1.0`
  as all negative values.
- Fix re-exporting `png::{CompressionType, FilterType}` to maintain SemVer
  compatibility with the `0.23` releases.

- Add ImageFormat::from_extension
- Add copyless DynamicImage to byte slice/vec conversion.
- Add bit-depth specific `into_` and `to_` DynamicImage conversion methods.

sha256sum:
7ce04077ead78e39ae8610ad26216aed811996b043d47beed5090db674f9e9b5  target/package/image-0.23.12.crate

Release 0.23.11

- The `NeuQuant` implementation is now supplied by `color_quant`. Use of the
  type defined by this library is discouraged.
- The `jpeg` decoder can now downscale images that are decoded by 1,2,4,8.
- Optimized the jpeg encoding ~5-15%.
- Deprecated the `clamp` function. Use `num-traits` instead.
- The ICO decoder now accepts an empty mask.
- Fixed an overflow in ICO mask decoding potentially leading to panic.
- Added `ImageOutputFormat` for `AVIF`
- Updated `tiff` to `0.6` with lzw performance improvements.

sha256sum:
b4f0a8345b33b082aedec2f4d7d4a926b845cee184cbe78b703413066564431b  target/package/image-0.23.11.crate

Release 0.23.10

- Added AVIF encoding capabilities using the `ravif` crate. Please note that
  the feature targets the latest stable compiler and is not enabled by default.
- Added `ImageBuffer::as_raw` to inspect the underlying container.
- Updated `gif` to `0.11` with large performance improvements.

sha256sum:
985fc06b1304d19c28d5c562ed78ef5316183f2b0053b46763a0b94862373c34  target/package/image-0.23.10.crate

Release notes for 0.23.10

Release 0.23.9

- Introduced correctly capitalized aliases for some scream case types
- Introduced `imageops::{vertical_gradient, horizontal_gradient}` for writing
  simple color gradients into an image.
- Sped up methods iterating over `Pixels`, `PixelsMut`, etc. by using exact
  chunks internally. This should auto-vectorize `ImageBuffer::from_pixel`.
- Adjusted `Clone` impls of iterators to not require a bound on the pixel.
- Add `Debug` impls for iterators where the pixel's channel implements it.
- Add comparison impls for `FilterType`

sha256sum:
974e194911d1f7efe3cd8a8f9db3b767e43536327e899e8bc9a12ef5711b74d2  target/package/image-0.23.9.crate

Release 0.23.8

- `flat::Error` now implements the standard `Error` trait
- The type parameter of `Map` has been relaxed to `?Sized`
- Added the `imageops::tile` function that repeats one image across another

Release 0.23.7

- Iterators over immutable pixels of `ImageBuffer` can now be cloned
- Added a `tga` encoder
- Added `ColorMap::lookup`, an optional reversal of the map
- The `EncodableLayout` trait is now exported

sha256sum:
a2397fc43bd5648b7117aabb3c5e62d0e62c194826ec77b0b4d0c41e62744635  target/package/image-0.23.7.crate