On the Predictive Accuracy of Neural Temporal Point Process Models for Continuous-time Event Data