ggml-cuda: fixes for concurrent streams#18496

am17an merged 4 commits intoggml-org:masterggml-org/llama.cpp:masterfrom

am17an:graph-opt-fixam17an/llama.cpp:graph-opt-fixCopy head branch name to clipboard