Cache-Augmented Generation (CAG) for HuggingFace models: encode your knowledge into a KV cache once, then answer questions without retrieval.