Two-Timescale Q-Learning with Function Approximation in Zero-Sum Stochastic Games