Self-attention in vision transformers performs perceptual grouping, not attention