Structure-agnostic Optimality of Doubly Robust Learning for Treatment Effect Estimation