llama.cpp

Implement non-mapped async IO for CUDA on Windows.

#7896

Merged