Robustness of Anomaly Detection Models for Industrial Control Systems under Training-Time Data Contamination