Data Attribution for Text-to-Image Models by Unlearning Synthesized Images