On Generalization of Graph Autoencoders with Adversarial Training