Split Adaptation for Pre-trained Vision Transformers