Boosting Large Language Models with Mask Fine-Tuning