Unleashing the Potential of Unsupervised Deep Outlier Detection through Automated Training Stopping