IntactKV: Improving Large Language Model Quantization by Keeping Pivot Tokens Intact