Robust Learning of Deep Time Series Anomaly Detection Models with Contaminated Training Data