Dual Grained Quantization: Efficient Fine-Grained Quantization for LLM