Removing Undesirable Concepts in Text-to-Image Generative Models with Learnable Prompts