[llvm] [AMDGPU] Add an option to run `AMDGPUAttributorPass` in full LTO (PR #102086)

[Johannes Doerfert via llvm-commits]

jdoerfert wrote:

We need to run the pass *once after linking* in all (supported) pipelines.
Running it multiple times is not helpful (as of now) and running it pre-linking with the runtime and other TUs is also not helpful.

Pipelines (for now):
OpenMP -> FullLTO
HIP w/ rdc
HIP w/o rdc
OpenCL

Pipelines (hopefully soon, https://github.com/jdoerfert/llvm-project/tree/thin_lto):
Thin-lto (OpenMP and HIP)
Partitioned-FullLTO

I think we need to move it late, as I said (and did) in the closed world patch after we moved it. Otherwise we miss out on closed-world *and* optimizations of runtimes and cross TU.

https://github.com/llvm/llvm-project/pull/102086