Update llama.cpp to fix MTP performance

Updating llama.cpp can yield a 1.5-1.8x token generation speed boost in MTP mode. Recent updates also fixed prompt processing performance issues.