Enhancing Ultra-Low-Bit Quantization of Large Language Models Through Saliency-Aware Partial Retraining