Learning Selection Bias and Group Importance: Differentiable Reparameterization for the Hypergeometric Distribution