ClusTR: Exploring Efficient Self-attention via Clustering for Vision Transformers