[llvm] [AMDGPU] Move kernarg preload logic to separate pass (PR #130434)

[Shilei Tian via llvm-commits]

shiltian wrote:

I think this is the best way to handle it. Moving it to the attributor isn't acceptable to me, at least based on what we have right now. I'd need to clearly see what benefits we get from running it iteratively, and how it would actually work if we were to make it part of the attributor.

https://github.com/llvm/llvm-project/pull/130434