Unbiased Gradient Estimation with Balanced Assignments for Mixtures of Experts