tokenizer:make-sentencepiece

(tokenizer:make-sentencepiece pieces scores &key specials bos eos)

Builds a SentencePiece-style BPE tokenizer -- Llama 2 and TinyLlama's, the one examples/llm/llm.lisp runs. pieces is a sequence of piece strings indexed by id and scores the matching merge scores; the encode adds a dummy prefix space, then repeatedly merges the adjacent pair whose concatenation scores HIGHEST. U+2581 is translated to a space, so a GGUF's tokenizer.ggml.tokens and a llama2.c tokenizer.bin's already-translated pieces are both accepted as they come, and the byte-fallback pieces <0x00>..<0xFF> are looked up by name rather than by position.