CoServe: Efficient Collaboration-of-Experts (CoE) Model Inference with Limited Memory