llama.cpp

vulkan: enable coopmat2 FA gqa and split_k optimizations more often

#12931

Merged