Relations, Negations, and Numbers: Looking for Logic in Generative Text-to-Image Models