LLMEmbed: Rethinking Lightweight LLM's Genuine Function in Text Classification