arxiv Towards Self-Interpretable Graph-Level Anomaly Detection