Generate Any Scene: Evaluating and Improving Text-to-Vision Generation with Scene Graph Programming