A Policy-Gradient Approach to Solving Imperfect-Information Games with Iterate Convergence