Fix merge_metadata for cases where the dimension of the variable changes

[mhauru]

The diff here is actually much simpler than it looks. The only real change in src/varinfo.jl is removing the line @assert length(vals_left) == length(vals_right), in the if vn in vns_left && vn in vns_right path. It's just that once you remove that line, you can simplify the surrounding code, because two of the if-paths become exactly identical and the third one nearly so.

I don't see any reason why we couldn't, when merging two metadata objects, merge just the same in the case where the dimension of some variable is different on the left and the right. I asked @torfjelde and he couldn't think of a reason either, but if someone else knows why the @assert was there in the first place, please say so.

This change would resolve TuringLang/Turing.jl#2472.

[coveralls]

Pull Request Test Coverage Report for Build 12885489364

Details

• 13 of 13 (100.0%) changed or added relevant lines in 1 file are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage decreased (-12.6%) to 73.721%

---

Totals
Change from base Build 12812100099:	-12.6%
Covered Lines:	3156
Relevant Lines:	4281

---

💛 - Coveralls

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 86.16%. Comparing base (938a69d) to head (0547f1b).
Report is 1 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #781      +/-   ##
==========================================
- Coverage   86.25%   86.16%   -0.10%     
==========================================
  Files          36       36              
  Lines        4330     4301      -29     
==========================================
- Hits         3735     3706      -29     
  Misses        595      595              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[coveralls]

Pull Request Test Coverage Report for Build 12907177100

Details

• 13 of 13 (100.0%) changed or added relevant lines in 1 file are covered.
• 3 unchanged lines in 1 file lost coverage.
• Overall coverage decreased (-0.09%) to 86.246%

---

Files with Coverage Reduction	New Missed Lines	%
src/varinfo.jl	3	83.87%

Totals
Change from base Build 12812100099:	-0.09%
Covered Lines:	3706
Relevant Lines:	4297

---

💛 - Coveralls

[penelopeysm]

Lgtm 👍

[mhauru]

Thanks @penelopeysm. I'll give this a few hours to see if anyone who knows the old code wants to rush in shouting "STOP THE MERGE".

[penelopeysm]

Weird. The only difference between the runs is that ADTypes was upgraded from 1.11.0 to 1.12.0, which was a change to AutoFiniteDiff that isn't used anywhere in our code.

[penelopeysm]

OK, I found that KernelAbstractions.jl also got bumped from 0.9.31 to 0.9.32, and I think it's this JuliaGPU/KernelAbstractions.jl#551 that broke our code.

It seems that for now a suitable solution is to pin the version of KernelAbstractions (we should do this in the toplevel Project.toml, not just the test one, because with the newer version JET will infer an untyped varinfo when it's not really needed). I tested locally and can confirm that this fixes the failing test.

Longer term we should make a smaller repro and post an upstream issue.

[penelopeysm]

ell gee tee em