perf(vision): fuse bilinear resize to grayscale

[rsasaki0109]

Summary

• add safe allocating and caller-owned fused bilinear RGB resize-to-gray APIs
• reuse Q11 sampling plans, preserve strided views and Gray metadata, and avoid the intermediate RGB image
• expose resize_rgb_to_gray_image(..., out=...) in Python with synchronized stubs and tests
• add Criterion coverage, a paired OpenCV 4.13 harness, README/roadmap/GitHub Pages updates, and a dated receipt

Results

On the recorded 12-thread Windows host, the allocated 1920x1080 to 960x540 pipeline measured 0.677 ms versus OpenCV's two-call 0.755 ms, a 1.12x SpatialRust win. Native caller-owned fused execution improves over SpatialRust's unfused path by 1.09x at 1080p, 1.15x at 4K, and 1.17x at 8K.

The claim is scoped: 4K allocated is effectively tied, while OpenCV leads 8K allocation and all caller-owned-output comparisons.

Fused output is bit-exact with the SpatialRust unfused plan. Canonical and 300 randomized OpenCV comparisons have maximum absolute error 1/255.

Validation

• cargo test -p spatialrust-vision --features full
• generated-size fused/unfused property test
• cargo clippy -p spatialrust-vision --features preprocess --all-targets -- -D warnings
• Python crate clippy
• Python binding pytest
• release maturin build
• focused paired OpenCV harness
• rustdoc, rustfmt, py_compile, and diff checks