Exploring Efficient Few-shot Adaptation for Vision Transformers