Reveal and Release: Iterative LLM Unlearning with Self-generated Data