CUDA: use mma FA kernel for gqa > 4 on RTX 4000#15035

JohannesGaessler merged 1 commit intoggml-org:masterggml-org/llama.cpp:masterfrom

JohannesGaessler:cuda-fa-kernel-choiceJohannesGaessler/llama.cpp:cuda-fa-kernel-choiceCopy head branch name to clipboard