LLMs Can Understand Encrypted Prompt: Towards Privacy-Computing Friendly Transformers