Towards Automated Self-Supervised Learning for Truly Unsupervised Graph Anomaly Detection