NativeAOT/win-x86: Enable FEATURE_EH_CALLFINALLY_THUNKS

[filipnavara]

Fixes #99687

[dotnet-policy-service]

Tagging subscribers to this area: @agocke, @MichalStrehovsky, @jkotas
See info in area-owners.md if you want to be subscribed.

[filipnavara]

cc @dotnet/jit-contrib

[filipnavara]

I run into some issues that may be related to this. For example, Invariants.Tests compiled without this change passes. With this change the stack pointer is trashed early in exception handling and causes ret to end up reading wrong location from stack after the catch funclet executed. I'm investigating.

[filipnavara]

Turns out the issue is unrelated to this change. It's simply lack of EECodeManager::GetResumeSp equivalent on the NativeAOT side (#99688 (comment)). That is a separate issue that needs to be fixed.

[AndyAyersMS]

@BruceForstall perhaps you should take a look?

[filipnavara]

This is hopefully the last missing piece for the win-x86 NativeAOT support.

[BruceForstall]

Looks ok to me.

Note that it is changing exception handling behavior of the existing Linux/x86 support. We don't build or test that (that I know of), and comments here indicate Linux/x86 is already broken for other reasons, so perhaps this adds yet another thing that either improves the existing state or adds to the backlog. In any case, it makes sense that x86 with funclets and x64 with funclets are on the same plan.