perf(vision): reuse morphology workspace

[rsasaki0109]

What changed

• add RectMorphologyWorkspace and caller-owned *_into APIs for rectangular u8 morphology
• retain full-image ping-pong planes and per-worker line buffers across calls
• expose Python MorphologyWorkspace and out= with shape, contiguity, and overlap validation
• benchmark allocated and reuse modes against OpenCV 4.13
• update README, roadmap, GitHub Pages morphology guide, and dated performance receipt

Why

The accelerated rectangular engine still rebuilt scratch planes and worker buffers on every public call. Explicit workspace/output reuse removes recurring host allocation cost while preserving safe APIs, exact output, explicit ownership, and generic-shape fallbacks.

Impact

On the recorded 12-thread Windows host, caller-owned 511×511 opening is 3.25× faster than OpenCV at 1080p and 2.77× faster at 4K. OpenCV remains faster for 5×5 rectangles, which is documented as the current optimization boundary.

Validation

• cargo test -p spatialrust-vision --features imgproc-morphology (14 passed)
• focused post-edit morphology tests (12 passed)
• cargo clippy -p spatialrust-vision --features imgproc-morphology -- -D warnings
• Python binding suite passed with one skip
• 980 randomized OpenCV parity cases
• Criterion benchmark target compiles
• rustdoc generation and staged diff checks pass