[SYCL][Fusion] Refine remapping of GEP instruction during internalization

[jopperm]

So far, we distinguished GEP instructions that select an element of the internalized buffer (must be remapped) and GEPs that address into an aggregate type (must not be remapped) by looking at the number of indices.

However, we can also encounter single-index GEP instructions that use a byte-offset to address into padded structures, as well as multi-index GEPs with a base pointer offset that address into an aggregate and need to be remapped.

This PR adds uses the newly added element size information (#12108) to correctly distinguish the required action for these kinds of GEP instructions. In addition to the E2E tests, target-specific lit tests derived from it are also added to demonstrate the subtle differences among SPIR-V, CUDA and HIP.

[jopperm]

Converted back to draft because of the CI failure on the newly-introduced test.

[victor-eds]

I'd rather include less tests targeting the cases we are handling here instead of .ll dumps. This would simplify testing and make sure we check all the boxes.

[sommerlukas]

Implementation looks fine, a few comments on the tests.

Overall, I do see value in having those tests also in .ll format to get an understanding what kind of input and transformation we expect in the compiler. However, maybe we can condense the input a bit by removing unnecessary stuff or running a bit of optimization on it first.

[jopperm]

I was able to condense the spir64 lit-test quite a bit, and will continue with the CUDA and HIP versions later. Reverted this PR back to draft until then.

[sommerlukas]

Tests look better now, but I think we can still improve readability.

In particular for metadata, we should only keep what we need for kernel fusion/the test and remove everything else.

[jopperm]

In particular for metadata, we should only keep what we need for kernel fusion/the test and remove everything else.

Agreed and done, I misinterpreted your earlier comment, sorry. I also shortened identifier names throughout the test cases.

[victor-eds]

Naming looks more sensible to me now