GPU-03 Launch path and GpuJob #22

Goal

Land the permanent submit and wait path that gpu() will use: bind a GpuPack, dispatch a compute pipeline, wait on a fence, write results back into Python.

Work

• Descriptor set layout for option 5: binding 0 = scalar SSBO, bindings 1..N = list SSBOs.
• Pipeline layout and compute pipeline created from SPIR-V. Build SPIR-V with the library (committed bytes or shaderc). Do not require glslangValidator on the user PATH.
• Command recording: barriers after upload if needed, bind descriptor set, dispatch, submit with a fence.
• GpuJob with join(): wait, download, write back into the same Python objects. No mid-run sync.
• Wire this to GpuPack from Issue 2. Same types later used by the public gpu() entry.
• End-to-end proof only through _ext.gpu.testing (same idea as pack round-trips). Keep any smoke kernel under testing. Do not add a public or product “reference kernel” API.
• Pytest: run the test launch, check results, skip when no GPU.

Notes

One compute queue. Results only after join. Everything here is permanent launch machinery, not a temporary demo path.