FBQuant: FeedBack Quantization for Large Language Models