Segment, Mask, and Predict: Augmenting Chinese Word Segmentation with Self-Supervision