MAPO: Boosting Large Language Model Performance with Model-Adaptive Prompt Optimization