Towards Optimal Off-Policy Evaluation for Reinforcement Learning with Marginalized Importance Sampling