Character, Word, or Both? Revisiting the Segmentation Granularity for Chinese Pre-trained Language Models