Learning Actionable Representations with Goal-Conditioned Policies