AugViT: Improving Vision Transformer Training by Marrying Attention and Data Augmentation