Revisiting Text-to-Image Evaluation with Gecko: On Metrics, Prompts, and Human Ratings