Towards Deep Attention in Graph Neural Networks: Problems and Remedies