Beyond Worst-case: A Probabilistic Analysis of Affine Policies in Dynamic Optimization