I lean towards this as the better solution (I started writing a comment about how I was even between them and in the process convinced my self to prefer this one).

This has the advantage of being cleaner and more localized but hard-codes the non-cooperative modules in. However, odds of another binding showing up are low and doing the sub-classing and replacing __init__ while waiting for changes to be merged upstream isn't that bad.

#9048 uses 'simpler' python, but it makes the class zoo one class bigger and has to use two differently named 'magic' cut-out function, and is spread across a couple of files. It makes the unlikely extension case easier (once you understand the code), but harder to understand to begin with.

try... finally added.
Honestly, I think anyone coming up with a new qt binding that does not support cooperative inheritance does not deserve being supported by matplotlib :-) (you can make the mistake once, but you should learn from previous mistakes!)

Appveyor failure is just a hang at the start of the conda-build on 3.6.