Fix tmerge for (partially) const types with undef fields

[martinholters]

When tmerge is applied to Const/PartialStruct and a field is Const in one of the types and Union{} (i.e. undef) in the other, the resulting field type must not be Const (as the resursively called tmerge produces).

Fixes #43784.

[martinholters]

Side note: Looks like the bug was exposed by #43305 which made getfield_tfunc return Union{} for undefined fields instead of their declared type. So it's probably not necessary to backport this patch.

[martinholters]

Ci failure on FreeBSD is unrelated Download failure ("Connection time-out while requesting https://httpbingo.julialang.org/base64/SnVsaWEgaXMgZ3JlYXQh").

[JeffBezanson]

Could we instead modify ⊑ to handle undefined fields better? It calls getfield directly, when it could call getfield_tfunc, and then it should just work.

[giordano]

I can confirm this fixes also #43561:

     Testing Running tests...
Test Summary: | Pass  Total  Time
Base          |    7      7  1.5s
Test Summary: | Pass  Total  Time
Utils         |    2      2  0.3s
Test Summary: | Pass  Total  Time
Promotion     |    6      6  3.4s
Test Summary: | Pass  Total  Time
Maths         |   18     18  1.3s
Test Summary: | Pass  Total  Time
Show          |    4      4  1.5s
     Testing PhysicalConstants tests passed 

julia> versioninfo()
Julia Version 1.8.0-DEV.1315
Commit 73d46513d1 (2022-01-14 11:20 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: Intel(R) Core(TM) i7-4870HQ CPU @ 2.50GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-13.0.0 (ORCJIT, haswell)

the fact that PhysicalConstants.jl hangs while precompiling (#43561) is due to the fact that Measurements.jl is broken (#43784)

[vtjnash]

Could we instead modify ⊑ to handle undefined fields better? It calls getfield directly, when it could call getfield_tfunc, and then it should just work.

You cannot. This PR is currently exploiting that flaw to achieve convergence here. When that is improved to "just work", we will need to also add this case to issimpleenoughtype, to prohibit complexity from accumulating here.

[martinholters]

CI failure is a timeout in the Distributed test which I'd think is unrelated. Good to go?

[aviatesk]

yep, let's go.