Token-Label Alignment for Vision Transformers