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