Probabilistic Inference in Reinforcement Learning Done Right