ggml-cuda: use passed ops instead of hardcoded ops#16712

am17an merged 1 commit intoggml-org:masterggml-org/llama.cpp:masterfrom

am17an:cuda-refactor-can-fuseam17an/llama.cpp:cuda-refactor-can-fuseCopy head branch name to clipboard