SecFormer: Towards Fast and Accurate Privacy-Preserving Inference for Large Language Models