llama.cpp

metal : do not use mul_mm kernels when ne00 < 64

#3542

Merged