Recover-to-Forget: Gradient Reconstruction from LoRA for Efficient LLM Unlearning