Eliminate recursion (#62)

## 0.11.0 2026-01-15

### Changed

* !Remove recursive implementations
  * All dimensionalities now use const-generic bounded implementations
* For dimensions higher than const-unrolling limit, a run-time loop of
the same form is used
  * No notable change in performance for high dimensions
  * No change to API for `::interpn` convenience functions
* All methods are now compatible with static analysis of memory usage
via cargo-call-stack
* Refactor internals to reduce duplication
* Improve form of early bounds checks to be interpreted correctly by the
compiler
  * 30% speedup in some cases, no effect on most
* Bounds checks done via iterator do not always result in optimizing out
panic branches
* Test cubic methods in higher dimensions

Implement thread parallelism (#61)

## 0.10.0 2026-01-10

### Added

* Rust
  * Add top-level `interpn`, `interpn_alloc` and `interpn_serial` methods along with supporting enums for selecting methods
  * Add `par` feature, enabled by default, that enables parallelism with rayon in `interpn` function
  * Add lazy static for getting number of physical cores to advise thread chunk sizes

### Changed

* Rust
  * Update deps
* Python
  * !Refactor `interpn` function
    * Combine `check_bounds: bool` and `bounds_check_atol: float` to `check_bounds_with_atol: float | None`
    * Replace `assume_regular` input with optional `grid_kind` to allow assuming either regular or rectilinear, or making no assumption
    * Add `max_threads: int | None` input to allow manually limiting parallelism
    * Use rust top-level `interpn` function as backend for method selection, bounds checks, and parallelism

Improve compile time (#59)

## 0.9.1 2025-12-31

### Changed

* Rust
  * Remove explicit const-unrolling for shallow loops (eliminates nested const-unrolling)
    * Improves compile time
    * Minimal effect on performance in Rust benchmarks
    * About 20% improvement for single-sample throughput from Python
* Python
  * Lock python package version to rust version
  * Improve handling of `__version__` when package is not present

Improve compile time (#58)

## 0.9.0 2025-12-27

### Added

* Rust
  * Add `deep-unroll` feature that sets crunchy unroll depth to 256 and enables 4D unrolled cubic interpolation
    * This improves compile times in typical use-cases

### Changed

* Rust
  * Gate 4D unrolled cubic interpolation behind `deep-unroll` feature
  * Enable `deep-unroll` feature for Python builds
  * Update pyo3 deps

Use x86-64-v3 reference CPU target (#52)

## 0.8.2 2025-11-12

### Changed

* Use `x86-64-v3` reference CPU as compilation target instead of manually specifying features
  * Includes AVX2 and vector FMA

Implement PGO for MacOS and Windows (#51)

## 0.8.1 2025-11-10

### Added

* Python
  * Implement PGO for mac intel/aarch64, linux aarch64, and windows x64
  * Add script for installing llvm on linux and mac to support PGO

### Changed

* Python
  * Remove stored pgo profiles
  * Update PGO scripts

Implement `interpn` convenience function (#47)

## 0.7.0 2025-10-25

Implement `interpn` convenience function that defers directly
to `raw` functions without serializable wrapper classes.

This enables use without pydantic for dep-constrained environments.

Contributions
* [Clément Robert](https://github.com/neutrinoceros)
  * #40
  * #41
  * #42
  * #44
  * #45
  * #46

### Added

* Python
  * Add type stubs for `raw` module
  * Add `interpn` function that defers directly to `raw` functions without serializable wrapper classes

### Changed

* Python
  * !Move pydantic dep to `serde` optional group
  * !Move optional dev dep groups to dependency groups (no longer installable via `.[dep]` syntax)
  * !Import serializable wrapper classes into top level module only if pydantic is found
  * !Drop support for python 3.9, which is leaving long-term support soon
  * Update PGO profile data
* Rust
  * Use abi3-py310 for python bindings

Nearest neighbor on structured grids (#43)

Nearest neighbor on structured grids

Fix typo in workflow (#39)

Use FMA (#29)

## 0.6.2 2025-10-20

Add optional use of fused multiply-add, enabled for python distributions.
This substantially improves floating-point roundoff; cubic method now shows
O(1e-14) peak roundoff error even under extrapolation of a quadratic function,
and 0-4 epsilon roundoff inside interpolating region.
Overall effect on throughput performance is neutral.

### Added

* Rust
  * Add `fma` feature

### Changed

* Rust
  * Use Horner's method for evaluating normalized cubic hermite spline
  * If `fma` feature is enabled, use FMA in cubic and linear methods where possible
* Python
  * Enable `fma` feature for python distribution
  * Update pgo profile data and benchmark plots