Faster prompt lookup drafting in llama.cpp

Researchers have achieved a 42x speedup in prompt lookup drafting for llama.cpp, a popular inference engine, through performance optimizations that also reduce memory usage by up to 2.6x. Prompt lookup decoding uses simple n-gram models to draft tokens faster during text generation, and the improvements optimize how llama.cpp manages its three types of n-gram caches (context, dynamic, and static). Further optimizations by Daniel Lemire brought the total speedup to approximately 140x.