Too Sure to Be Safe: Model Calibration for Reliable Log Anomaly Detection