When Do Supervised UQ Ensembles Improve LLM Hallucination Detection? A Robustness Study