Learning Optimal Power Flow: Worst-Case Guarantees for Neural Networks