Revisiting the Performance of Deep Learning-Based Vulnerability Detection on Realistic Datasets