Revisiting Graph Contrastive Learning on Anomaly Detection: A Structural Imbalance Perspective