Quantile-constrained Wasserstein projections for robust interpretability of numerical and machine learning models