Bridging the Gap for Tokenizer-Free Language Models