Accelerate direct and fused Sobel

[rsasaki0109]

What

• add exact direct first-order 3x3 Sobel APIs for grayscale u8 to f32
• add fused saturated-absolute and binary-threshold Sobel APIs with caller-owned outputs
• borrow packed NumPy inputs without copying and explicitly pack strided inputs
• add randomized OpenCV comparison coverage, Criterion cases, Pages documentation, and a dated receipt

Why

The canonical CPU table showed standalone Sobel trailing OpenCV by 20.31x to 23.30x at large resolutions. The generic path allocated a full-image f64 separable intermediate, while the Python binding also copied packed input.

Impact

• exact CV_32F-compatible Sobel X is 1.88x faster at 1080p and 2.03x faster at 4K; VGA remains an OpenCV 1.07x win
• fused abs(Sobel X) > 96 masks are 3.81x to 6.64x faster in allocate mode and 2.95x to 8.68x faster with caller-owned output
• 300 randomized X/Y cases and all canonical outputs are bit-exact
• no unsafe; generic Sobel remains the fallback for other kernels/orders/types

Checks

• cargo test -p spatialrust-vision --features imgproc-filter
• cargo clippy -p spatialrust-vision --features imgproc-filter --benches -- -D warnings
• cargo check --manifest-path crates/spatialrust-py/Cargo.toml
• .venv/Scripts/python.exe -m pytest crates/spatialrust-py/tests/test_bindings.py -q -k "advanced_filters or spatial_gradient or sobel"
• .venv/Scripts/python.exe bench/opencv_sobel_threshold_comparison/performance.py --output target/opencv-sobel-threshold-performance.json
• .venv/Scripts/python.exe bench/opencv_vision_comparison/performance.py --profiles vga,1080p,4k --output target/opencv-comparison/vision-performance-sobel-direct.json