perf(vision): beat OpenCV on detection NMS

[rsasaki0109]

What changed

• cache validated box areas across greedy NMS, batched NMS, and Soft-NMS overlap checks
• borrow contiguous NumPy score arrays and retain the non-contiguous packing fallback
• add native Criterion and schema-valid OpenCV Python NMS performance harnesses
• publish Epic 114G results in README, ROADMAP, receipt, and GitHub Pages algorithm catalog

Why

NMS repeatedly recomputed both box areas for every candidate/selected comparison. Detection outputs commonly contain hundreds to 8,400 candidates, so invariant geometry dominated post-processing.

Correctness

• ordered kept indices exactly match OpenCV dnn.NMSBoxes for deterministic 100, 1,000, and 8,400 candidate profiles
• cached-area IoU is bit-identical to the public IoU implementation across overlap, disjoint, identical, and zero-area cases
• contiguous and non-contiguous NumPy score arrays follow the same result path

Performance receipt

Windows 11, CPython 3.12.10, OpenCV 4.10, OpenCL off, randomized interleaved Python calls:

• 100 candidates: SpatialRust 8.95x
• 1,000 candidates: SpatialRust 3.82x
• 8,400 candidates: SpatialRust 3.22x

Validation

• spatialrust-vision detection tests: 5 passed
• spatialrust-py binding tests: 68 passed, 1 skipped
• OpenCV report contract tests: 8 passed
• vision and Python crate Clippy with warnings denied
• native Criterion detection benchmark
• changed Rust file rustfmt check and git diff check