LLM Unlearning Should Be Form-Independent