llama.cpp

CUDA: batched+noncont MMQ, refactor bs>1 MoE code

#13199

Merged