Test against newer Sphinx versions

[befeleme]

Hey,
As suggested in #277 I've added newer Sphinx versions to the CI. It uncovers some issues with Sphinx 3.5 and 4.0, which may be worth looking into.

[codecov]

Codecov Report

Merging #279 (b3c6aaf) into master (928573f) will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master     #279   +/-   ##
=======================================
  Coverage   98.90%   98.90%           
=======================================
  Files           3        3           
  Lines         273      273           
  Branches       46       46           
=======================================
  Hits          270      270           
  Partials        3        3           

[pquentin]

Thank you @befeleme! 🎉 This is very useful. Now, as you say, someone needs to look into this, eg. by adding commit to this PR or creating a new one based on this one.

I'd prefer not merging this as is unless the CI is green. One option is to use pytest's xfail in strict mode to encode that Sphinx 3.5 and 4.0 currently break.

[njsmith]

Looks like it's just our fragile tests failing due to irrelevant changes in sphinx's output again. It would be nice if we had a better way to test the rendered output than these fragile regexes :-(

Maybe there's a better way to write the regexes? Maybe we should switch to another strategy like pytest-snapshot? maybe we should just grit our teeth and update the regexes?