fix(lint): clippy::float_cmp + redundant cast (CI -D warnings)

CI's `cargo clippy --workspace --all-targets -- -D warnings` (clippy 1.95)
failed two lints that local dev-profile clippy didn't escalate:

1. butteraugli/src/lib.rs:1131 — `assert_eq!(res.max_norm(), 1.234)` triggers
   `clippy::float_cmp`. Switch to `to_bits()` comparison: max_norm() is just
   a getter, must round-trip exactly.

2. butteraugli-bench/examples/score_parity.rs:208 — `rust_score as f64` is
   a redundant cast (rust_score is already f64).

release: butteraugli 0.9.1 — dep bumps (archmage 0.9.14)

chore: bump archmage/magetypes minimum to 0.9.5, cargo update all

Merge pull request #2 from imazen/release/0.8.0

release: butteraugli 0.8.0 — cleanup, SIMD migration, CI expansion

release: bump to 0.7.2 — NEON + WASM SIMD128 support

Add ARM64 NEON and WASM SIMD128 SIMD dispatch for all hot-path
functions (blur, diff, malta). Cross-architecture parity verified
against hash-locked x86_64 reference scores (191 test cases).

release: bump to 0.7.1

- Fix cross-platform compilation (macOS ARM64, WASM)
- Eliminate planar→interleaved→planar round-trip in compare_linear_planar
- Update to Rust edition 2024
- Remove ARM64 NEON row from Performance table (scalar fallback only)
- Replace retired windows-2019 CI runner with windows-2022

release: bump to 0.7.0 — remove MaltaVariant (breaking API change)

release: bump to 0.6.2 — clippy warning fixes

release: bump to 0.6.1 — updated README with verified parity numbers

Validated against libjxl butteraugli_main on 21 real photograph pairs:
- 10 images at Q75 (576x576): max 0.0003% relative difference
- 6 images at Q50/Q90 (576x576): max 0.0001%
- 5 large images at Q75 (1024-2048px): max 0.0001%

Also fixed cargo install instructions (butteraugli-cli crate).

fix: exclude examples/ from crate package