How to Train your Text-to-Image Model: Evaluating Design Choices for Synthetic Training Captions