Validate GPU spatial dispatch before allocation

[rsasaki0109]

GPU nearest resize could reach texture creation with unsupported dimensions, and Sobel could initialize pipelines before rejecting a foreign runtime. Both could allocate work before profiler-capacity rejection. Validate runtime ownership, dimensions, existing compute/binding limits and profiler capacity before allocation/cache initialization; also reject overflow in resize's shader coordinate products.

Valid shader math, public signatures, metadata, one-pass receipts and explicit transfers are unchanged. Shared-halo Sobel remains a test-only experiment: initial gains did not reproduce consistently across repeat measurements, including adverse large/odd-strip cases. No production speedup is claimed.

Validation:

• Five new tests: 54 Sobel fixtures / 14,400,648 exact RGBA bytes against independent CPU and frozen shaders, plus 42 resize mappings / 104,888 exact RGBA bytes. Invalid limits/capacity/ownership, strided ROI, concurrent cold calls, live outputs and pool reuse are covered.
• 142 native GPU/image/point tests pass; strict default/all-feature Clippy, two WASM compilation profiles, formatting and diff checks pass. WASM execution is not claimed.
• The production shader, frozen fixture and shared executor match the parent exactly. The final 36-condition release diagnostic checks every output and publishes all paired comparisons. Its frozen shader uses the current host executor, so this does not measure historical API overhead.

Implementation and full measurement note.

Based on merged PR #143 (main edaf620). Merge is gated on the frozen source head, all 114 named checks succeeding, no pending/failed checks and a clean merge state.