Ward: Provable RAG Dataset Inference via LLM Watermarks