Reasoning about Counterfactuals to Improve Human Inverse Reinforcement Learning