Mixed-Precision Graph Neural Quantization for Low Bit Large Language Models