[clang] [llvm] [Offload][CUDA] Allow CUDA kernels to use LLVM/Offload (PR #94549)

[Johannes Doerfert via cfe-commits]

jdoerfert wrote:

> Should the NFCI changes (like initializing struct fields) be put into a separate PR?

Done.

https://github.com/llvm/llvm-project/pull/95116 
https://github.com/llvm/llvm-project/pull/95117

> Will kernels in TUs compiled with -foffload-via-llvm be interoperable with code that wants to launch them from another TU compiled w/o -foffload-via-llvm ?

I don't think they are/reasonably can be, at least not for now. I added a test to show 2 TUs with `-foffload-via-llvm` work together as expected though. Do you think we want such a setup to work? Either, we can follow up on this.


https://github.com/llvm/llvm-project/pull/94549