Add resident GPU Gaussian image filtering

[rsasaki0109]

Gaussian blur was available only on CPU, requiring image readback before weighted filtering in an otherwise resident GPU image pipeline.

Add gaussian_blur_gpu and the optional two-timestamp profiled variant under gpu-image. GpuGaussianBlurOptions supports independent 3/5/7 kernels, finite positive sigmas, all 1..=4 logical channels and five borders. The host prepares two tiny CPU-compatible coefficient vectors; the two weighted image passes execute on GPU in one submission. Q8/Q15 kernels, unrounded horizontal sums and two-u32-limb Q15 final sums preserve the existing specialized CPU u8 bytes exactly.

Returned images retain metadata and independent ownership and compose with gray/Sobel/AI operations. Private scratch uses 4*channels bytes per pixel; output writes use physical RGBA8. Transfers remain explicit, with no image payload readback, repeated image upload or host wait inside the filter. Dimensions, arithmetic, scratch/uniform/binding/workgroup capacities and optional profiler ownership/capacity reject before resource/cache allocation. Unsupported scratch sizes return an error; placement remains caller-controlled.

Validation:

• Six native Gaussian tests: 3,240 exact CPU conditions / 3,176,550 logical bytes, including asymmetric kernels/sigmas, strips, all borders and logical channels.
• Extreme sigmas, constant-255 Q15 mass and unused RGBA components; a 1025x1025 grid-stride case; exact reduced device/uniform/scratch bounds; invalid runtime/options/profiler allocation checks; ROI metadata, retained/concurrent output and dirty scratch reuse.
• Complete GPU-image library regression: 79 passed, 13 intentional ignored measurements. Facade Gaussian/resident gray-edge-AI smoke, Clippy all features/targets, formatting, feature-free GPU and GPU-image WASM builds passed.
• Two isolated release surveys, 12 cases each, four stages, three warmups and 20 rotating paired rounds with exact output checks. Compared with reused parallel CPU workspace/output, resident 7x7 1080p grayscale is 3.21-4.23x faster and 4K RGB is 2.03-2.26x faster. Upload/readback reduces gains; light grayscale cases remain faster on CPU. Full tables and within-survey intervals retain every adverse result.

A private CPU resize helper now compiles only for its actual resize/test consumers, fixing the dead-code lint when only filter/analysis reference features are enabled. No production dependency, core model, CPU Gaussian behavior or unsafe boundary changes. Existing Vision-chain defaults remain.

Evidence: C:/Users/rsasa/Workspace/SpatialRust/notes/2026-10-13_gpu_gaussian_blur.md. Architecture, ROADMAP and CHANGELOG describe the API and arithmetic/transfer boundaries. Follows merged PR #178.