Unify NMSs and FastPreProcess as FastVisionOps

[Som5ra]

Summary

This PR combines NMSs and FastPreProcess into FastVisionOps, one validated CPU toolkit for vision inference preprocessing and postprocessing.

• adds a primary fastvisionops package while preserving nmss imports
• adds validated HWC/CHW conversion and fused channel normalization
• unifies native preprocessing and bounding-box NMS in one rebuildable C library
• uses OpenMP when available with a portable compiler fallback
• fixes the two self-review edge cases for empty mask IoU and empty NMS validation
• adds 41 correctness and native-equivalence tests
• adds reproducible elapsed-time benchmarks and CI smoke tests
• rewrites the README and evaluation report with migration guidance for both repositories

Correctness

• 41/41 unit tests pass
• native NMS matches NumPy across 128 randomized parameter combinations
• native preprocessing matches NumPy for randomized single and batched inputs
• noncontiguous images, empty batches, RGB/BGR reversal, invalid statistics, and thread validation are covered
• serial and concurrent batch NMS output matches item by item
• the C source compiles with -Wall -Wextra -Werror
• the source wheel builds and contains the unified C source

Measured performance

Each value is the median elapsed time for one complete public API call after two warm-up runs and nine measured runs. Validation, allocation, and array preparation are included. Native output is checked against NumPy before timing.

Fused preprocessing, 427x640x3 images, 8 native threads

Batch	NumPy	Native	Speedup
1	4.492 ms	0.249 ms	18.07x
8	26.263 ms	1.888 ms	13.91x
32	139.105 ms	9.758 ms	14.25x

Bounding-box NMS

Boxes	NumPy	Native	Speedup
250	4.798 ms	0.272 ms	17.66x
1,000	23.257 ms	3.340 ms	6.96x
2,500	75.396 ms	17.121 ms	4.40x

Eight images with 1,000 boxes each took 27.595 ms with serial native execution and 9.995 ms with eight workers, a 2.76x speedup.

Environment: Linux x86_64, Python 3.12.13, NumPy 2.3.5, GCC 13.3, 9 available Intel Xeon Platinum 8573C vCPUs.

Validation commands

python -m fastvisionops.build
python -m unittest discover -s tests -v
python -m benchmarks.benchmark_preprocess
python -m benchmarks.benchmark_bbox
python -m pip wheel --no-build-isolation --no-deps .

Commit sequence

1. fix: handle empty NMS API inputs
2. feat: introduce unified FastVisionOps package
3. feat: unify native vision operations
4. perf: add preprocessing acceleration benchmark
5. docs: present the unified FastVisionOps toolkit
6. fix: validate execution limits consistently
7. docs: refresh final validation evidence
8. fix: validate empty native NMS batches
9. docs: tighten FastVisionOps overview

The earlier focused NMS commits remain intact in the branch history. Commit messages contain no emoji.

[Som5ra]

Self-review completed. GitHub does not permit the PR author to request changes or approve their own PR. Two reproducible API edge cases remain. Core native correctness, concurrency stress, packaging, and all CI jobs passed.

[Som5ra]

Final self-review completed with no unresolved code findings. Verified 40/40 tests, randomized NumPy/native equivalence, strict C warnings, OpenMP and portable fallback builds, source-wheel packaging, benchmark correctness checks, and clean review threads. GitHub does not allow a pull request author to approve their own PR, so this comment records the successful self-review result.

[Som5ra]

Independent sub-agent review result: APPROVE. The reviewer confirmed the empty native-batch validation fix, 41/41 tests, the full Python 3.9/3.12/3.13 CI matrix, and the concise 173-line README. No bugs, logical errors, or documentation findings remain. GitHub rejected a formal APPROVE because the connected account is the pull request author, so this comment records the approval recommendation.