test: remove broken marker, better define notimpl and notyet

[NickCrews]

inspired by #9535

The broken mark had the meaning of "this broke for some reason, but it wasn't the fault of this PR, so lets ignore it for now". I find this only meaningful in the context of that individual PR though: as soon as the PR is merged, the usefulness drops to near 0. I come across this mark as I'm reading code and I have no idea what action I should take. notimpl, notyet, and never are much more useful, they signal where the fault lies and what to do about it.

I went through and converted all the instances of broken to one of these other 3 marks. It was fairly fast and definitely error prone. I could see the argument for converting these all to plain xfails when we aren't sure. I'm not sure how bad it is for these marks to be incorrect.

I tweaked the prose that describes these 3 marks. I would love your editing of them.

[gforsyth]

Thanks for putting this in @NickCrews -- I agree the utility of broken vanishes after a PR is merged.

At a glance, the marks you've chosen look reasonable - some of them may be wrong, but it's both subjective and imperfect, so I think we can just deal with it.

I think we should try this for a bit -- we may end up restoring something similar. To your point about "plain xfails", that may be the path we want to take (with whatever extra stuff we'll need to specify which backends to xfail for a given test).