Support Atomic(T)#compare_and_set when T is a reference union

[HertzDevil]

Atomic(T)#compare_and_set was not working when T is a union of reference types, because it wrongly used the code path for nilable reference unions. This PR fixes that by removing legacy code that is only relevant before LLVM 3.9.

It also adds checks against T in the arithmetic methods to make sure T is an arithmetic type rather than a pointer. LLVM would produce module validation errors otherwise if Atomic doesn't catch those errors.

[Sija]

I'd suggest adding a NOTE admonition.

ditto elsewhere

[straight-shoota]

Can you elaborate why? This seems like a very basic part of the description, rather than a callout.

[Sija]

It's a hard requirement, thus IMO deserves visibility.

[straight-shoota]

But that's not exactly what NOTE admonitions are for.

[oprypin]

It's hard to use admonitions or agree on what they should mean. I think NOTE: should be the same as a footnote, but it's hard to stick to this definition because the reality is that the current look increases visibility rather than decrease it.

[Sija]

@straight-shoota Oh, I didn't know they're defined (not in the docs they aren't at least) ;)

[HertzDevil]

It turns out LLVM doesn't support pointers in #swap until 15.0: llvm/llvm-project@18e6b82

[beta-ziliani]

Would it be sensible to still add support for it for LLVM >= 15?

[HertzDevil]

We just need to bring back the existing pointerof(@value).as(LibC::SizeT*) workaround in #swap. I am more surprised by all the years it took them