Norm Tweaking: High-performance Low-bit Quantization of Large Language Models