PLA: Prompt Learning Attack against Text-to-Image Generative Models