Centroid Transformers: Learning to Abstract with Attention