Faster prompt lookup drafting in llama.cpp

Four changes to the n-gram caches of llama.cpp make drafting for prompt lookup decoding up to 41.6x faster per drafted token, load the static cache up to 23.5x faster, and lower peak memory up to 2.65x.