SOUL: Unlocking the Power of Second-Order Optimization for LLM Unlearning