A Statistical Learning Perspective on Semi-dual Adversarial Neural Optimal Transport Solvers