Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt