The Causal Structure of Domain Invariant Supervised Representation Learning