Subtype: avoid some false alarm in subtype_unionall

[N5N3]

The current check is not correct if we set the typevar's bounds to an unwrapped UnionAll. Use var_occurs_inside_skip to skip the inside check.
close #24333, close #47654

[vtjnash]

I will try to review and understand this at some point. In the meantime, I thought you might be interested in #47813 and #47877 also, since they also deal with subtyping/intersection.

[vtjnash]

@nanosoldier runtests(ALL, vs=":master")

[nanosoldier]

Your package evaluation job has completed - possible new issues were detected. A full report can be found here.

[N5N3]

Failure in Unitful shows that b5f7982 fails to pick the right env.
So perhaps we'd better avoid the UnionAll unwrapping when we meet Union with typevar.

[N5N3]

@nanosoldier runtests(ALL, vs=":master")

[nanosoldier]

Your package evaluation job has completed - possible new issues were detected. A full report can be found here.

[N5N3]

The result looks good. The only concern is MosekTools, its test passed on my PC with a local solver, but I saw some random error if I use the network fallback solver.

Since the "Segmentation fault" was happened in libmosek64, I thought this PR might be unrelated? cc @blegat @ulfworsoe

[blegat]

I would assume it is not related to this PR Indeed, @ulfworsoe might know better where this crash is coming from, see https://s3.amazonaws.com/julialang-reports/nanosoldier/pkgeval/by_hash/2bac38a_vs_b6f32bc/MosekTools.primary.log

[vtjnash]

As I look further into this, I think the concept of var_occurs_inside and especially in_union is somewhat flawed. I made a counter-example of this here: #31167 (comment)

My current theory is that we need to instead deal with this by setting the var in the env to something unsatisfyable (either with a bit, or with a falsifiable constraint such as Any <: ub <: Union{}) whenever computing the intersection for the lb and ub

[ulfworsoe]

@blegat It is possible that the crash is solely a MOSEK issue. I don't see how it would be related to a anything on the Julia side, unless it's a memory corruption. We have also seen segfaults when allocation fails, e.g. if a memory limit was set with ulimit.

[JeffBezanson]

I may have a fix for #47658 that also fixes these cases.