perf(vision): fuse resize normalization and CHW packing

[rsasaki0109]

Summary

• add safe allocating and caller-owned APIs that fuse Q11 bilinear RGB resize, float normalization, and planar CHW packing
• avoid the intermediate resized HWC image while preserving reusable plan semantics, strided input, metadata, and arbitrary scale/mean/std
• expose resize_normalize_image_chw(..., out=...) in Python with synchronized stubs and tests
• add native Criterion coverage, a paired OpenCV blobFromImage harness, README/roadmap/GitHub Pages updates, and a dated receipt
• complete Epic 115C

Results

Against OpenCV 4.13 cv2.dnn.blobFromImage on the recorded 12-thread Windows host, SpatialRust allocated calls measured:

• 1080p to 640x640: 1.617 ms vs 3.570 ms, 2.21x faster
• 4K to 640x640: 2.117 ms vs 4.272 ms, 2.02x faster
• 4K to 1280x720: 3.592 ms vs 8.359 ms, 2.33x faster

Caller-owned SpatialRust output is 3.02x-3.56x faster than OpenCV allocation. Native fused reuse improves over SpatialRust unfused reuse by 1.10x-1.15x.

Fused output is bit-exact with SpatialRust's unfused plan for arbitrary normalization parameters. Three hundred randomized OpenCV comparisons have maximum float error 0.003921628 (one uint8 level after 1/255 scaling).

Validation

• cargo test -p spatialrust-vision --features full (124 unit, 13 property)
• vision and Python clippy
• Python binding pytest
• release maturin build
• paired OpenCV harness with 300 randomized cases
• rustdoc, rustfmt, py_compile, and diff checks