The Struggles of Feature-Based Explanations: Shapley Values vs. Minimal Sufficient Subsets