Unified Parameter-Efficient Unlearning for LLMs