On the Sample Complexity of Batch Reinforcement Learning with Policy-Induced Data