Testing Relational Understanding in Text-Guided Image Generation