GPU compute probe - Vulkan backend, one kernel end to end

The GPU workstream's first milestone: obSCEne dispatches a compute shader and reads the result bits back (D109). Proven on llvmpipe in the build VM - plumbing correct - and the identical code runs on a Deck's RDNA2 for real.

Landed: gpu.h interface; gpu_vulkan.c (host/Deck) and gpu_gnm.c (refusing console stub), source-list split; src/shaders/*.comp + scripts/gen_shaders.py embedding SPIR-V into a committed header; 160-gpu section with a pipeline self-check (exact x*2, asserted) and a rcp observation (emits input/output bits per lane); gpudev/gpu records with device provenance; GPU=1 build flag. All four build combinations compile clean, verify: ok.

Surprises

• A standalone Vulkan spike ran first and confirmed the VM can dispatch compute on llvmpipe before any integration - the whole GPU direction rested on that being possible, and it is.
• counts.py counts check rows as text, so the first gpu.c with two #if/#else check tables would have counted three checks. Restructured to one table with the bodies varying under the flag - stable count, cleaner diffs.
• Vulkan is verbose under the project's -Wconversion -Wsign-conversion -Werror, and a /* inside a generated comment (src/shaders/*) failed -Wcomment. Both are the strict gate doing its job.