llama.cpp runs CUDA, ROCm together

This is the kind of setup that makes llama.cpp feel more like a backend router than a single-GPU runtime, but it is still a power-user build, not a turnkey feature.

• –`GGML_BACKEND_DL` is the key enabler: backends can load dynamically, so one executable can target different GPU stacks.
• –The example output shows a practical hybrid split across CUDA, ROCm, and host memory, which matters when a model no longer fits cleanly on one vendor's card.
• –The Windows toolchain is brittle enough that `GGML_CPU_ALL_VARIANTS=ON` needed manual pruning, so reproducibility will depend on exact compiler and driver versions.
• –For local-LLM tinkerers with mismatched GPUs, this is a real win; for production, the maintenance burden likely outweighs the gains unless the hardware mix is fixed.