PromptGuard: Soft Prompt-Guided Unsafe Content Moderation for Text-to-Image Models