Exploiting Hierarchy for Learning and Transfer in KL-regularized RL