Towards Confidential and Efficient LLM Inference with Dual Privacy Protection