Validate GPU AI packing before allocation

[rsasaki0109]

GPU CHW image packing could acquire output storage before detecting unsupported output-binding/device limits or insufficient profile capacity. Validate source dimensions/channels, u32 pixel/element indices, checked output bytes, existing buffer/compute/binding limits and profiler ownership/capacity before allocation or pipeline setup.

The production shader, pipeline factory, CHW ordering, f32 normalization and tensor ownership/readback remain unchanged. No new scratch, payload transfer, pipeline, dependency or public signature is introduced. The two attempted optimizations stay native-test-only: channel lanes regressed across all selected GPU conditions; fixed-channel micro gains did not establish completed-call improvement and included adverse cases. No speedup is claimed.

Validation:

• Six new tests verify 23,857,560 exact output words across 192 channel/shape/normalization cases against frozen GPU words, including independent exact CPU cases, both experimental shaders, extreme finite parameters, ROI/concurrency/live tensor reuse and invalid limits/ownership/capacity.
• A 4096x4096 resident Vision chain verifies 16,777,216 zero tensor words, six passes, unchanged payload receipts and 96 separately collected query bytes. Oversize declared-source rejection uses a metadata-only negative probe rather than a giant image allocation.
• 154 native image/point GPU tests pass; strict default/all-feature Clippy, two WASM compilation profiles, formatting and diff checks pass. Browser execution is not claimed.
• All 24 final resident timing controls are published, with two warmups and twenty alternating rounds each. The frozen shader shares the current host executor, so this is not historical API-overhead measurement. Pilot results and adverse samples are retained in the note.

Implementation and measurement note.

Based on merged PR #146 (main a9ba90e). Merge is gated on the frozen source head, all 114 named checks succeeding, no pending/failed checks and a clean merge state.