LLM Unlearning via Loss Adjustment with Only Forget Data