Update CUDA binaries with mmq fix (b8190 -> b8191)

Docker Space: llama.cpp CUDA inference with multi-GPU load balancing

initial commit