What MI-based Representation Learning Objectives are Sufficient for Control?