Training Design for Text-to-Image Models: Lessons from Ablations