A Dual Approach to Constrained Markov Decision Processes with Entropy Regularization