Forgetting Only What Matters: Layer-Selective Unlearning toward Robust LLMs