Duplicated values need to be sorted according to weight as well as mean in MergingDigest

[tdunning]

If there are multiple centroids with the same mean we should prefer to sort the ones with higher weight towards the median.

The reason for doing this is that without this tweak on the sort, centroids with the same mean that previously for which the size constraint was satisfied could be inverted during the sort and, as a result, now violate the constraint. That leads to reduced accuracy and apparent non-deterministic behavior. It can also cause problems because the first and last centroid can wind up with weights greater than one. This can adversely affect accuracy since the interpolation algorithms assume that the extreme centroids have unit weight.

This problem is known to occur in the Merging Digest, but it is likely to be true of the AVLTreeDigest as well. This issue only has to do with the MergingDigest. Issue #155 deals with the corresponding problem in the AVLTreeDigest.

[tdunning]

This was closed by eca1125
Instead of fancy sorting by weight, I changed the sort to be stable.