perf(vision): accelerate bilinear resize

[rsasaki0109]

Summary

• add reusable BilinearResizeU8Plan with precomputed Q11 half-pixel coordinates and coefficients
• add an exact row-parallel packed RGB8 half-scale path and route Python bilinear resize through it
• preserve packed/strided input and output, metadata, padding, and Python out= identity
• add Criterion coverage, a focused OpenCV harness, dated receipt, README table, ROADMAP, and GitHub Pages updates

Measured results

• Native reuse vs previous SpatialRust: 35.6x VGA, 47.8x 1080p, 37.3x 4K
• VGA 640x480 -> 320x240 Python reuse: SpatialRust 0.120 ms vs OpenCV 0.133 ms (1.10x faster)
• OpenCV still leads reuse at 1080p/4K/8K by 2.40x/2.01x/1.85x
• Canonical half-scale output is bit-exact; 300 arbitrary-dimension cases have maximum absolute error 1

Validation

• 9 focused resize tests, including strided input/output and plan validation
• 118 full vision unit tests + 11 property tests
• 8 spatialrust-image tests
• vision all-target clippy with -D warnings
• Python binding suite and Python crate clippy (excluding one pre-existing argument-count lint)
• release extension build, rustfmt, rustdoc, focused OpenCV harness, and diff checks