A Theory of Regularized Markov Decision Processes