[Quixyl-eng]

Summary

• Add exvram-cuda, a Windows launcher that starts the CUDA llama-server build with the same full-GPU flags the benchmarks already use.
• The decoder stays on CUDA0. --embedding-gpu on pins the token embedding there, matching the VRAM planner. The server stops when the process exits.

Test plan

• Built with Visual Studio 2022 Build Tools (/W4, no warnings) on the RTX 5060 checkout.
• One prompt run printed gpu_layers 25/25, CUDA0 model buffer 1211.05 MiB, graph_splits 1, ggml_cuda_loaded yes, and the answer 4. llama-server was gone afterward.
• Confirm a machine without the local GGUF and runtime fails before starting a server, as the README describes.