Leveraging Programmatically Generated Synthetic Data for Differentially Private Diffusion Training