Compatibility: pyarrow 0.20dev causes test failures

[bsipocz]

We need to look into the regressions more, it is very likely just output related like the case we have for mpl. In that's true, we may need to pin its version in the test matrix and bump it in a more controlled fashion when we also bump the references.

https://github.com/executablebooks/MyST-NB/pull/659