Do-GOOD: Towards Distribution Shift Evaluation for Pre-Trained Visual Document Understanding Models