llama.cpp

CUDA/HIP: Share the same unified memory allocation logic.

#12934

Merged