Breaking the Sample Complexity Barrier to Regret-Optimal Model-Free Reinforcement Learning