Test llama.cpp threads argument for up to 80% performance boost

A user reports up to 80% performance improvement by tuning the --threads argument in llama.cpp for hybrid CPU architectures. The tip suggests using only P-cores with taskset/affinity for optimal inference speed.