Evaluating Logical Generalization in Graph Neural Networks