Exploring the effects of adversarial data and model robustness on quantifying epistemic uncertainty for chest radiographs