LBPE: Long-token-first Tokenization to Improve Large Language Models