Lizard: An Efficient Linearization Framework for Large Language Models