LARA: A Light and Anti-overfitting Retraining Approach for Unsupervised Anomaly Detection