perf(vision): accelerate rectangular morphology

[rsasaki0109]

What changed

• add a safe separable prefix/suffix min/max engine for packed grayscale u8 rectangular morphology
• preserve the generic fallback for other shapes, types, channels, strides, anchors, and borders
• dispatch contiguous NumPy rectangles through the fast path and preserve non-contiguous input behavior
• add Rust/Python parity tests, Criterion profiles, and a focused interleaved OpenCV harness
• update the README comparison table, Epic 117 roadmap state, dated receipt, and GitHub Pages algorithm guide

Why

The previous implementation scanned every active kernel sample for every output pixel, leaving 5×5 opening hundreds of times slower than OpenCV. Rectangles are separable, so the new path computes each line with block prefix/suffix extrema and bounded parallel row/column stages.

Measured impact

• 4K 5×5 opening: 4599.72 ms → 105.05 ms (43.8× faster than the prior SpatialRust path)
• 1080p 511×511 opening: SpatialRust 22.559 ms vs OpenCV 55.543 ms (2.46×)
• 4K 511×511 opening: SpatialRust 100.889 ms vs OpenCV 207.788 ms (2.06×)
• small-window scope is reported honestly: OpenCV remains faster at 5×5

All published cases are bit-exact. The harness also passes 980 seeded randomized OpenCV comparisons across all seven operations.

Validation

• cargo test -p spatialrust-vision --features imgproc-morphology
• cargo clippy -p spatialrust-vision --features imgproc-morphology -- -D warnings
• cargo bench -p spatialrust-vision --bench morphology --features imgproc-morphology --no-run
• python -m pytest crates/spatialrust-py/tests/test_bindings.py -q (68 passed, 1 skipped)
• python bench/opencv_morphology_comparison/performance.py --output target/opencv-morphology-performance.json