Long-Context Inference with Retrieval-Augmented Speculative Decoding