Breaking the Sample Size Barrier in Model-Based Reinforcement Learning