On the Variance of Unbiased Online Recurrent Optimization