On the Optimality of Sparse Model-Based Planning for Markov Decision Processes