Negative Preference Optimization: From Catastrophic Collapse to Effective Unlearning