perf(vision): accelerate Gaussian blur

[rsasaki0109]

Summary

• add safe gaussian_blur_u8 and reusable gaussian_blur_u8_into APIs with GaussianBlurU8Workspace
• accelerate RGB8 3x3/5x5 Gaussian blur with cached Q8 kernels, border/interior specialization, target-feature dispatch, and row parallelism
• keep 7x7 on the established high-precision path; add NumPy out= support, tests, Criterion coverage, an OpenCV comparison harness, and GitHub Pages documentation

Measured results

• Native 5x5 allocated path: 20.7x faster at 1080p, 26.7x faster at 4K than the previous SpatialRust implementation
• Native caller-output + workspace path: 48.4x / 65.9x faster than the previous allocated path at 1080p / 4K
• OpenCV remains faster for standalone Python calls: 3.10x / 2.93x / 3.58x at 1080p / 4K / 8K
• Canonical 5x5 profile is bit-exact; 300 randomized 3x3/5x5/7x7 cases have maximum absolute error 2

Validation

• 17 focused Rust filter tests
• 115 full vision unit tests + 11 property tests
• clippy with -D warnings
• release Python extension build and full binding suite
• Criterion Gaussian benchmark group
• 300-case OpenCV accuracy/performance harness
• rustfmt, rustdoc, and staged diff checks