RoSTE: An Efficient Quantization-Aware Supervised Fine-Tuning Approach for Large Language Models