LUNAR: LLM Unlearning via Neural Activation Redirection