Accelerating LLaMA Inference by Enabling Intermediate Layer Decoding via Instruction Tuning with LITE