[llvm] [AMDGPU][Passes] Invoke CGSCCOptimizerLateEP callbacks in LTO pipeline (PR #203262)

[Shilei Tian via llvm-commits]

shiltian wrote:

> Full LTO is still a TU in the strictest sense, is the concern compile times?

I guess it's easy to have a CGSCC in the per-TU case, but it's probably not worth it once you have the full graph, except for calls into external libraries, which don't really matter anyway.

https://github.com/llvm/llvm-project/pull/203262