Adversarial Training for Graph Neural Networks: Pitfalls, Solutions, and New Directions