GWQ: Gradient-Aware Weight Quantization for Large Language Models