Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning