UnStar: Unlearning with Self-Taught Anti-Sample Reasoning for LLMs