I don't really like breaking encapsulation, but you can't get insight
about the statistical performance of a t-digest without iterating over
its centroids.