llama.cpp

cmake: revert CUDA arch default to 52, 61 if f16

#1959

Merged