Cross-model Control: Improving Multiple Large Language Models in One-time Training