Fork of llama.cpp adds --numa mirror mode for multi-socket CPU inference

A fork of llama.cpp adds a --numa mirror mode to optimize inference on multi-socket CPU systems. The author is seeking testers for the new branch.