llama.cpp

CUDA: use mma FA kernel for gqa > 4 on RTX 4000

#15035

Merged