Collaborative Speculative Inference for Efficient LLM Inference Serving