LICHEE: Improving Language Model Pre-training with Multi-grained Tokenization