[llvm] [AMDGPU] Add AMDGPU-specific module splitting (PR #89245)

[Pierre van Houtryve via llvm-commits]

Pierre-vh wrote:

> I found a bug during the review of #89683. I will attach the test files here. These tests should pass when they are run individually and fail when they are tested as a group. These nondeterministic results show up when `AMDGPUSplitModule` or `SplitModule` are used for partitioning. I suspect it may have something to do with tie breaks during load balancing.
> 
> [clone-lds-function.ll.txt](https://github.com/llvm/llvm-project/files/15202789/clone-lds-function.ll.txt) [clone-lds-function-ancestor-kernels.ll.txt](https://github.com/llvm/llvm-project/files/15202790/clone-lds-function-ancestor-kernels.ll.txt) [clone-lds-function-successor.ll.txt](https://github.com/llvm/llvm-project/files/15202791/clone-lds-function-successor.ll.txt)

I'm not able to reproduce it, can you create a testcase that only depends on this pass?
I had to modify the tests a lot to make them work so it possibly lost what makes them non-deterministic. Also I see you're using `%u` but I don't know what that's supposed to do, I replaced it with %t to make it work.

https://github.com/llvm/llvm-project/pull/89245