Polynomial Convergence of Gradient Descent for Training One-Hidden-Layer Neural Networks