Training Normalizing Flows with the Precision-Recall Divergence