Beyond additivity: sparse isotonic shapley regression toward nonlinear explainability