VIREL: A Variational Inference Framework for Reinforcement Learning