A self-supervised world model that jointly learns two architecturally distinct representations: one equivariant to specified transformations and another invariant to them