[AMDGPU] Use reverse iteration in CGP

[Pierre-vh]

#64591 has been fixed, so once the dust settles a bit (at least until the change propagates to downstream branches and we don't have any major issues) we can move forward with:

• Removing all of the i16 promotion logic from CGP, deleting the option and removing/reworking associated tests
• Use a reverse iteration order in the pass so it works like a normal combiner.

[llvmbot]

@llvm/issue-subscribers-backend-amdgpu

Author: Pierre van Houtryve (Pierre-vh)

#64591 has been fixed, so once the dust settles a bit (at least until the change propagates to downstream branches and we don't have any major issues) we can move forward with:

• Removing all of the i16 promotion logic from CGP, deleting the option and removing/reworking associated tests
• Use a reverse iteration order in the pass so it works like a normal combiner.