LMEraser: Large Model Unlearning through Adaptive Prompt Tuning