Add Canny magnitude ring

[rsasaki0109]

What changed

• replace the full-frame Canny comparison-magnitude buffer with three rows per worker
• preserve parallel row-stripe NMS and exact hysteresis
• skip hysteresis traversal when classification finds no weak edges
• expose reusable workspace allocated bytes in Rust/Python and focused receipts
• update README and GitHub Pages results

Why

The remaining full i32 magnitude image consumed 31.6 MiB at 4K and added avoidable memory traffic. High-contrast inputs also paid for graph traversal even when every retained edge was already strong.

Impact

The fast path stays bit-exact with OpenCV across 300 randomized cases. Native 4K document lines improve from 96.914 ms with inspectable intermediates to 8.134 ms with ring reuse (11.92×). Python reuse measures 8.103 ms versus OpenCV 11.480 ms (SpatialRust 1.42×); 1080p is 1.36× faster. Dense sensor noise remains an OpenCV win and is documented as such.

Validation

• cargo test -p spatialrust-vision --features imgproc-canny (26 passed)
• targeted Clippy with -D warnings
• Python crate check and Canny binding tests (2 passed)
• focused OpenCV 4.13 harness, 300 randomized bit-exact cases
• Criterion 4K document-line receipt