Reduce resident GPU Canny gradient memory by half

[rsasaki0109]

Pack resident GPU Canny's bounded signed derivatives into one u32 and retain a separate u32 magnitude, reducing each gradient from 16 to 8 bytes without changing edges or public options. A 4K call saves 66,355,200 private scratch bytes; scratch becomes 20N+8 (aperture 3) or 28N (5/7). Planning and destination receipts reflect the smaller binding, with the same explicit transfers and five/six passes.

Validation: 864 exact CPU/frozen-GPU/compact-GPU edge conditions, independent gradient/NMS/BFS snapshots, all 65,536 signed component values and 326,401 raw rounding values, extremal kernels, capacity limits, ROI, profiling, concurrency and resident chain reuse. GPU-image regression: 87 passed, 15 explicit measurements ignored. GPU all-feature/all-target Clippy, facade chain, feature-free GPU, GPU-image WASM and formatting checks passed.

Two isolated paired release surveys compare the actual frozen PR #180 shader with the compact shader. At 4K, resident completed-call ratios are 1.22–1.84x and full upload/readback ratios 1.04–1.12x under default 64 MiB pool budgets. Smaller full calls show no established speed gain, and resident results vary. The measurements include resource acquisition and completion; they do not isolate shader throughput or imply automatic CPU/GPU placement. See notes/2026-10-13_gpu_canny_compact_gradients.md for all samples' summaries, intervals and conditions.