VPTQ: Extreme Low-bit Vector Post-Training Quantization for Large Language Models