Robust Offline Policy Evaluation and Optimization with Heavy-Tailed Rewards