Reversing the Forget-Retain Objectives: An Efficient LLM Unlearning Framework from Logit Difference