CUDA: use registers instead of smem in topk-moe#16647

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

am17an:cuda_topk_optimizeam17an/llama.cpp:cuda_topk_optimizeCopy head branch name to clipboard