On the Fairness, Diversity and Reliability of Text-to-Image Generative Models