ggml-cuda: enable cuda-graphs for n-cpu-moe#18934

am17an merged 3 commits intoggml-org:masterggml-org/llama.cpp:masterfrom

am17an:n-cpu-moe-piecewiseam17an/llama.cpp:n-cpu-moe-piecewiseCopy head branch name to clipboard