TokenMix: Rethinking Image Mixing for Data Augmentation in Vision Transformers