cuda: software-FP16 fallback for fattn dequantize_V<half> on CC < 600#3

cchuter merged 1 commit into

feat/v4-port-cudacchuter/llama.cpp:feat/v4-port-cudafrom

fix/fattn-half-dequantizecchuter/llama.cpp:fix/fattn-half-dequantizeCopy head branch name to clipboard