Evaluating Structural Generalization in Neural Machine Translation