kmeans test is flaky due to custom make-blobs being random

[seberg]

The kmeans test is still somewhat flaky and sometimes fails. We can reduce the number of clusters, dimensions, etc. to ensure better convergence. Maybe sorting the centers may also help, but not sure it is likely that two centers swap in the first dimension. (the current code sorts the first dimension to match centers, rather than finding the best matches. That is because the center order itself is random).

Alternatively, and maybe better may be to see if we can't just stability the random-ness. And one easy way to do so to not use the custom make-blobs, but instead use sklearns one.