CITER CITER: Collaborative Inference for Efficient Large Language Model Decoding with Token-Level Routing