Uncertainty-Aware Subset Selection for Robust Visual Explainability under Distribution Shifts