Use normalize_options() from Docutils in preference to set_classes()

[bsipocz]

Purpose

We started seeing deprecations for set_classes() when running sphinx-automodapi tests (e.g. astropy/sphinx-automodapi#214). We don't call that method directly, but it appeared that sphinx itself calls it. Using this patch locally, those failures cleaned up.

However, this patch really badly fails the sphinx tests locally, so I'm trying to smoke out issues with CI (and I expect that some docutils version checker will be needed, too. I'll add changelog/authors once this is all green and decided to be the good approach)

[AA-Turner]

We still need the old name for old versions of Docutils, which we still support

[bsipocz]

OK, I suspected that this needs to be version dependent. I'll add a checker, though first will need to sort out what goes wrong for the latest docutils where it should work

[bsipocz]

Ahh, and this annoys Ruff, though I explicitly added the line as it was very difficult to see where the signature ends and the actual code starts.

[bsipocz]

The one remaining failure is actually relevant as instead of the <h2 class="myclass rubric">A rubric with a heading level 2</h2> we now generate <h2 class="rubric">A rubric with a heading level 2</h2> while in the input we clearly have :class: myclass.

I suspect this is something trivial for someone more familiar with the sphinx testing internals, so I don't yet dive into this rabbit hole of figuring it out myself.