llama.cpp

1f8fa523 - metal: add the 128/96 flash attention kernels, accept every compiled pair (#30209)