Distributionally Robust Learning with Weakly Convex Losses: Convergence Rates and Finite-Sample Guarantees