Improving the Robustness of Representation Misdirection for Large Language Model Unlearning