llama.cpp

9c72e7e9 - rebase to master (except ggml-cuda)