Gradient play in stochastic games: stationary points, convergence, and sample complexity