Add allocation-light Canny path

[rsasaki0109]

What changed

• make ordinary 3×3 canny() avoid public intermediate images
• add safe strided canny_into and reusable CannyWorkspace
• add Python out= / workspace support and a focused OpenCV harness
• refresh the README CPU vision table and GitHub Pages algorithm catalog with measured results

Why

The previous implementation always materialized gradients, magnitude, and suppression images even when callers only requested the final edge map. That allocation and conversion overhead made the published OpenCV comparison stale after the newer reusable CPU work.

Impact

The focused OpenCV 4.13 receipt is bit-exact across 300 randomized cases. The standalone OpenCV lead is reduced from the superseded 10.66×–12.65× table row to 1.65×–1.77× on reusable document-line profiles. OpenCV still leads, and Epic 118D remains open.

Validation

• cargo test -p spatialrust-vision --features imgproc-canny (25 passed)
• cargo clippy -p spatialrust-vision --features imgproc-canny -- -D warnings
• cargo check --manifest-path crates/spatialrust-py/Cargo.toml
• Python binding Canny tests (2 passed)
• focused VGA/1080p/4K harness, 300 randomized bit-exact cases