Offline Policy Evaluation and Optimization under Confounding