Add exact Euclidean distance transform

[rsasaki0109]

What changed

• add safe exact Euclidean distance transforms with unit and anisotropic pixel spacing
• expose distance_transform_edt to NumPy, accepting conventional zero/nonzero masks
• add known-grid, invalid-contract, exhaustive property, Python, and OpenCV precise-L2 tests
• add native Criterion and paired OpenCV VGA/1080p/4K performance coverage
• publish the algorithm and honest local accuracy/performance receipt in README, Pages, and notes

Why

SpatialRust had binary masks, contours, connected components, and morphology but no exact distance field. EDT is a reusable foundation for mask thickness, watershed seeds, contour distance, feather blending, and signed-distance workflows.

The implementation follows Felzenszwalb and Huttenlocher's separable lower-envelope algorithm and runs in O(width × height). A non-empty all-foreground mask is rejected because it has no finite nearest-background distance.

Measured impact

• OpenCV correctness maximum error: 9.536743e-7 on the irregular case
• canonical VGA/1080p/4K masks: exact values
• native Criterion: 12.585 ms / 100.98 ms / 451.63 ms
• Python median: 19.795 ms / 157.172 ms / 640.121 ms
• OpenCV remains 10.60× / 12.63× / 12.35× faster on the recorded Windows host

Validation

• cargo test -p spatialrust-vision --features full — 95 unit + 11 property tests
• cargo test -p spatialrust-platform — 10 tests
• cargo clippy -p spatialrust-vision --features dense --tests --bench dense -- -D warnings
• cargo doc -p spatialrust-vision --features full --no-deps
• cargo bench -p spatialrust-vision --features dense --bench dense -- --noplot
• Python binding suite — 67 passed, 1 optional skip
• mypy stubtest — pass
• OpenCV correctness and performance reports — pass
• comparison report contract — 8 tests

Workspace full Clippy still encounters the pre-existing spatialrust-camera manual-div-ceil lint before reaching this crate; the changed dense feature passes warnings-as-errors independently.