perf(vision): accelerate Soft-NMS beyond OpenCV

[rsasaki0109]

What changed

• replace Soft-NMS repeated sorting/removal with a one-pass active-candidate decay, compaction, and next-best scan
• short-circuit disjoint IoU work and borrow contiguous Python score arrays without an unconditional copy
• add exact sorting-reference regressions, non-contiguous Python coverage, and Criterion baselines
• add a reproducible OpenCV comparison harness and publish results in the README, roadmap, algorithm catalog, vision page, and measurement note

Why

Soft-NMS is useful for crowded detections because it decays overlapping scores instead of removing every overlap. The existing implementation preserved the algorithm but repeatedly sorted and shifted the candidate vector, adding avoidable overhead.

Measured impact

On the documented Windows 11 / Intel 6-core test host, SpatialRust is 3.42x to 7.40x faster than OpenCV 4.10 across linear and Gaussian Soft-NMS workloads from 100 to 8,400 boxes. Selected indices match exactly; maximum updated-score error is 1.79e-7.

Validation

• cargo test -p spatialrust-vision --features detection (8 passed)
• cargo clippy -p spatialrust-vision --all-targets --features detection -- -D warnings
• cargo clippy --manifest-path crates/spatialrust-py/Cargo.toml --all-targets -- -D warnings
• Python binding tests (67 passed, 1 skipped)
• OpenCV report contract tests (8 passed)
• changed Rust files pass rustfmt --check
• cargo doc -p spatialrust-vision --features detection --no-deps

Workspace-wide rustfmt still reports the repository's pre-existing formatting baseline in unrelated files such as stereo.rs; this slice does not modify those files.