Add resident GPU Canny edge extraction

[rsasaki0109]

CPU Canny required returning resident image data to the host before gradient, suppression and weak/strong edge connectivity.

Add canny_gpu and its optional profiled variant under gpu-image. Signed integer Sobel gradients support 3/5/7 apertures; exact threshold conversion, L1/squared-L2 comparisons, aperture-7 ties-to-even scaling and directional NMS ties match CPU semantics. Weak candidates form complete eight-neighbor components using monotone atomic union/path halving. Each weak component is flagged when it touches a strong pixel; strong pixels survive directly. This preserves full hysteresis without a propagation cutoff or host convergence/readback.

Five/six passes share one submission and return an owned binary GpuImage with preserved dimensions/metadata. Input remains borrowed; grayscale conversion and blur remain explicit caller stages. Source/output/runtime ownership, private submitted scratch, buffer/binding/workgroup and profiler checks are validated before resource/cache allocation. Uniform controls are 96 bytes; no image/label readback, repeated image upload, CPU fallback or automatic placement occurs. Receipt documentation clarifies per-stage destination footprints versus repeated atomic byte traffic.

Validation:

• 864 exact CPU conditions / 367,056 edge bytes across apertures, norms, threshold boundaries/reversal/extremes, strips and four profiles.
• 24 owned f64-gradient/NMS/BFS snapshots, including forced grid strides; 18 graph masks with complete roots/flags/output checks, long 1025x1025 weak snakes, far strong seed and a broken path.
• All 326,401 possible signed aperture-7 raw derivative values match independent binary64 ties-to-even rounding.
• Invalid limits/runtime/options/profiler allocation guards, strided ROI metadata, four concurrent retained results, dirty scratch, timestamp receipts and Gaussian/Canny/AI resident chaining.
• GPU image regression: 85 passed / 14 intentionally ignored measurement fixtures. Facade smoke, Clippy all features/targets, formatting, feature-free GPU and GPU-image WASM checks passed; final weak-only connectivity also passed all six Canny tests.
• Two isolated release surveys: 12 cases, four stages, three warmups and 20 rotating paired rounds, all output-byte exact. Completed upload/filter/readback calls for apertures 5/7 are 4.98-7.55x current CPU API speed at 1080p and 6.60-13.55x at 4K. Aperture-3 completed calls are slower on GPU, and resident gains are not universal. The CPU 5/7 into path still allocates inspectable intermediates; large GPU memory/time spikes and all adverse samples are documented.

Private scratch is 28N+8 bytes (3) or 36N (5/7), with a largest 16*N binding. Memory/device limits remain explicit; no production dependency, core model, CPU algorithm or unsafe boundary changes. Existing Vision-chain defaults remain.

Evidence: C:/Users/rsasa/Workspace/SpatialRust/notes/2026-10-13_gpu_canny.md. Architecture, ROADMAP and CHANGELOG describe the final weak-component algorithm, output/transfer contract and measured tradeoffs. Follows merged PR #179.