Multilingual BPE tokenizer family (128K/64K/8K) for LUNA-1B, with benchmark vs the original GPT-NeoX tokenizer.