Multi-manifold Attention for Vision Transformers