CUDA: fix padding of GQA to power of 2 in FA#19115

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

JohannesGaessler:cuda-fa-fix-gqa-paddingJohannesGaessler/llama.cpp:cuda-fa-fix-gqa-paddingCopy head branch name to clipboard