Structuring Representation Geometry with Rotationally Equivariant Contrastive Learning