SFT-GO: Supervised Fine-Tuning with Group Optimization for Large Language Models