[llvm] [AMDGPU] Shrink S_MOV_B64 to S_MOV_B32 during rematerialization (PR #184333)

[Romanov Vlad via llvm-commits]

romanovvlad wrote:

> I think this is continuing the bad trend of hacking in more contextual work into this function, which should just perform the mechanical changes required. The context doing the rematerialzation should know all the information about which lanes are needed. e.g., this function should pass in the live lane mask

Initially didn't want to modify non AMDGPU specific part. Now updated to propagate LaneBitmask.

https://github.com/llvm/llvm-project/pull/184333