A Unified Causal View of Domain Invariant Representation Learning