When Does Synthetic Data Augmentation Improve Score-Based Imbalanced Classification?