Use ReshapedArray for isbitsunion arrays (fixes #28611)

[timholy]

The only non-obvious part of this is getting correct dispatch for Array{Any}, ensuring that we return an Array without messing up inference of other cases. (It fails

julia/test/core.jl

Lines 4172 to 4174 in f4c7779

	arrayset_unknown_dim(Any, 1)
	arrayset_unknown_dim(Any, 2)
	arrayset_unknown_dim(Any, 3)

if we don't do this properly.)

Fixes #28611.

CC @GlenHertz, @nalimilan.

[timholy]

The failure on apple seems spurious.

[nalimilan]

Thanks! But don't we need a more general fix handling all unions of isbits types?

[timholy]

Hmm, good point. I had forgotten this was no longer special-cased for Missing.

[timholy]

Technically this is breaking, although the cases it breaks are already silently buggy. If the breaking nature is a problem, perhaps the backport label should be removed.

[timholy]

I'll merge soon, barring further concerns.

[KristofferC]

This doesn't seem backportable.

[andreasnoack]

What about the bug in #28611

[KristofferC]

It's in the same spot as those reffed by this PR: #29406.

[timholy]

What about the bug in #28611

That's what this fixes (see the test). I'm not sure I understand the relevance of #29406.

[KristofferC]

I think Andreas point is that this should be backported because it fixes a bug. My response is that even though it fixes a bug it might break working code (which is the same as the bugs getting fixed by #29406). AFAIU, we are quite conservative when it comes to behavior changes for patch releases.