Towards Generalizable Graph Contrastive Learning: An Information Theory Perspective