Desiderata for Representation Learning: A Causal Perspective