perf(vision): accelerate class-aware batched NMS

[rsasaki0109]

What changed

• bucket accepted batched-NMS indices by class so candidates never scan kept boxes from unrelated classes
• expose a one-call spatialrust.batched_nms NumPy API with contiguous borrowing and non-contiguous fallbacks
• add Rust reference-parity tests, Python binding contracts, and native Criterion coverage
• add an exact-index OpenCV dnn.NMSBoxesBatched comparison harness and register it in the benchmark manifest
• publish the measured outcome in README, ROADMAP, GitHub Pages, and a dated receipt

Why

Class-aware NMS is a canonical object-detection post-processing step. The old Rust core preserved the correct semantics but scanned every globally kept detection and discarded unrelated classes inside the hot loop. Keeping a vector per class removes those comparisons without changing the global score-ordered result.

Measured impact

Windows 11, 6-core/12-thread Intel CPU, CPython 3.12.10, OpenCV 4.10, OpenCL off; seeded interleaved Python API medians:

Profile	OpenCV	SpatialRust	Speedup
1,000 candidates / 20 classes	3.5377 ms	0.1341 ms	26.38×
8,400 candidates / 80 classes	211.7618 ms	2.1776 ms	97.25×

Both profiles return exactly the same ordered kept indices. Another 32 randomized OpenCV parity cases also passed locally.

Validation

• cargo test -p spatialrust-vision --features detection
• cargo clippy -p spatialrust-vision --all-targets --features detection -- -D warnings
• cargo clippy --manifest-path crates/spatialrust-py/Cargo.toml --all-targets -- -D warnings
• changed Rust files pass rustfmt --check
• full Python binding suite passes (one existing skip)
• OpenCV report-contract tests and generated report validation pass
• native Criterion: about 98.3 µs at 1,000 and 2.42 ms at 8,400 on a separately seeded 80-class workload
• cargo doc -p spatialrust-vision --features detection --no-deps
• git diff --check