On the Fundamental Impossibility of Hallucination Control in Large Language Models