Fix LLVM IR expectations for the autodiff abi_handling tests

[sgasho]

We had two stale expectations in the LLVM IR checks.

• L61: noundef is no longer emitted
• L80: return type has changed

Also, I made all noundef checks optional for test robustness. I thought noundef is less important than types and args for this test.

r? @ZuseZ4

[sgasho]

I tested this PR still works after enzyme submodule update: #157716

[ZuseZ4]

I think I've seen some of these changes before, so my concern is that either we continue to make the tests increasingly weaker, or we'll keep swapping back and forth between different outputs.
In the past, I added the noPostOpt flag, which made it a little better, but since the output still changed, we should find a more robust solution before putting it in CI.
Can you have a look at the -C passes=val -- a list of extra LLVM passes to run (space separated) flag, as well as some of the enzyme tests which run extra opts after enzyme?
https://github.com/EnzymeAD/Enzyme/blob/c96508349d9f121957292308c646b4a1fd53c507/enzyme/test/Enzyme/ReverseMode/add.ll#L2

We probably want to remove -Zautodiff=NoPostopt and replace it with -Zautodiff=Passes=... or something like that.
There is a -Zoffload=Host=path flag, which is also nested as an example.
The idea is that not setting the flag would result in the current behavior (full O3 after Enzyme), setting it to "" would be equal to the current NoPostOpt flag (which can then be removed) and by setting anything else we can specify the actual passes to run, similar to Enzyme.