[SYCL][Fusion] Adapt internalization to opaque pointers.

[jopperm]

Fixes sycl/test-e2e/KernelFusion/internalize_array_wrapper.cpp which was temporarily deactivated in #9828.

Switching to opaque pointers meant that accesses to the data structure in the testcase were no longer represented as one "deep" GEP, but rather a series of GEPs into the different nesting levels of the structure. Rather than bailing out, we now keep track of GEPs that index into an aggregate object, and exclude their users from the modulo-LocalSize rewriting (the pointer types still have to be changed to the desired target address space, though).

In addition, perform -simplifycfg before internalization in order to get rid of unreachable code which, in this particular case, contained Phi nodes on the candidate accessor.

[sommerlukas]

Do we need to add a LIT test checking that we do not promote a pointer into an aggregate?

[victor-eds]

Looks good. Can you add some lit tests to sycl-fusion/test/internalization/ testing new cases?

[jopperm]

CI failure looks unrelated. I'm still missing a testcase for aborting propagation into function calls, though.

[jopperm]

I've added a lit test for stopping the promotion of calls in InAggregate mode, and changed the corresponding error message to an NYI (rather than "it's unsafe") -- I think it could be supported, but we'd need to change the implementation to propagate the InAggregate flag on a per-argument-basis, which I'd argue is out of scope for this PR.

[jopperm]

Failure on CUDA is caused by a test added by this PR; investigating.