Adversarial Nibbler: A Data-Centric Challenge for Improving the Safety of Text-to-Image Models