Tim Peters: "Victor, happy to add comments, but only if there's
sufficient interest in actually using this."

It looks like tests fail on some platforms. I guess that it depends on
the different factors: quality of the C math library, quality of the
IEEE 754 hardware implementation, etc.

I like the idea of providing an accurate and *portable* function
(_nth_root) thanks to the usage of the decimal module to get a better
precision.

If someone considers that the performance of
statistics.geometric_mean() matters, I suggest to document it. It's
new feature of Python 3.6, it's not like a regression, so it's ok.

What do you think?