Downstream Task-Oriented Generative Model Selections on Synthetic Data Training for Fraud Detection Models