Add CPU image decoders with measured per-format backends

[MilkClouds]

Image decoding now has a TorchCodec-style CPU API without a Pillow or PyTorch runtime dependency. Images use dedicated codec backends selected through pixel comparisons and release-build measurements instead of routing through FFmpeg.

• Add decode_image, decode_jpeg/png/webp/gif/avif/heic, and ImageReadMode, returning NumPy CHW/NCHW arrays. Support mode/dtype conversion, orientation, alpha, animations, JPEG lists/CMYK, PNG 8/16-bit samples, and high-depth AVIF/HEIC.
• Use Rust png for PNG; libjpeg-turbo for JPEG; libwebp/libwebpdemux for WebP; giflib for GIF; and libavif/dav1d/libyuv for AVIF. Wheel builds bundle these dependencies with their licenses. HEIC continues to use optional system libheif. Audio/video retain FFmpeg.
• Add textured/progressive/subsampled image regressions, PNG palette/transparency/Adam7 and mode/dtype parity, malformed-input checks, AVIF alpha/sequences, and Pillow-free wheel runtime tests. Animated PNG remains explicitly unsupported.
• Include a reproducible comparison against TorchCodec, Pillow, image and libjpeg-turbo-rs. Across 54 measured RGB inputs, the selected implementation matched TorchCodec pixels exactly; JPEG/WebP latency was roughly comparable and PNG was about 1.7x faster. This is a limited x86 workload, not a general performance or full-format parity guarantee. Experimental crates are isolated from production dependencies.

Validation:

• pytest --compare: 444 passed locally.
• Ruff lint/format, Cargo fmt and Clippy with -D warnings passed.
• Source-built image dependencies and a repaired Linux wheel were validated locally. A fresh uv environment with only NumPy and the wheel decoded image/animation, audio and video fixtures with no Pillow, TorchCodec, PyTorch or PyAV.
• Local repaired wheel: 13.28 MiB download / 33.09 MiB unpacked, within existing size limits. Local ABI tag is manylinux_2_34; portable manylinux2014 remains the existing release CI build target.

See docs/image-backends.md for versions, methodology, raw reports, alternatives and limitations. Python runtime dependencies remain NumPy only. This PR does not add average_rate or change the released version.

[MilkClouds]

Superseded by a smaller Python-level OpenCV image adapter. The replacement PR will retain the comparison findings without adding native image builds or vendored codec sources.