Rethinking and Scaling Up Graph Contrastive Learning: An Extremely Efficient Approach with Group Discrimination