[llvm] [NFC][MISched][AMDGPU] Expose tracePick to be used in AMDGPU backend (PR #184664)

[via llvm-commits]

lijinpei-amd wrote:

> > Reasoning should be given when closing PRs
> 
> The pr tries to fix 3 minor points in GCNSchedStrategy:
> 
> 1. tracePick is not called as in the general MachineScheduler.
> 2. do-while on condition that should be asserted. This is already changed in MachineScheduler.
> 3. Candidate reason not accurate.
> 
> As these points are minor, I think it is not worthwhile to pursuit it.

4. MachineScheduler use ScheduleDAGMILive to refer to the pre-RA scheduler, which tracks register liveness, while ScheduleDAGMI to refer to the post-RA scheduler, which doesn't track register liveness.  GCNSchedStrategy use GCNPostScheduleDAGMILive to refer to the post-RA scheduler, in which `Post` and `Live` contradicts.

https://github.com/llvm/llvm-project/pull/184664