Simplex Neural Population Learning: Any-Mixture Bayes-Optimality in Symmetric Zero-sum Games