When Linear Attention Meets Autoregressive Decoding: Towards More Effective and Efficient Linearized Large Language Models