The Impact of Data Distribution on Q-learning with Function Approximation