llama.cpp src/llama-vocab.cpp (master) - token_to_piece and the per-token piece cache

Holds cache_token_to_piece, the cache of token id -> text piece, i.e. the table that token_to_piece() consults on every grammar accept and every detokenisation. Establishes that the per-token host cost has a lookup against a vocab-sized cache plus the piece bytes, not a re-tokenisation of the whole sequence.