Upstream baseline df03399b8 (b10902), cuda-12.9 recipe

Unmodified upstream llama.cpp commit df03399b8 (build b10902), compiled with the cuda-12.9 recipe of this repository's release workflow (CUDA 12.9.1, SASS sm_70/86/89/120a/121a, PTX sm_50/61/70/75/80/90), no downstream patches.

It is the upstream side of the benchmark grids in benches/multi-gpu/ and is fetched by the bench image (scripts/multigpu/bench/vast/). Not a release of llama.cpp-multigpu; for those see the multigpu-* releases. Needs the CUDA 12.9 runtime libraries (cudart-… archive of any multigpu-* release, or NVIDIA's nvidia/cuda:12.9.1-runtime image) and an R525+ driver (R575+ for PTX-only GPUs).