Unlearning of Knowledge Graph Embedding via Preference Optimization