Zero-Shot Sharpness-Aware Quantization for Pre-trained Language Models