Provably Efficient Neural Estimation of Structural Equation Models: An Adversarial Approach