AMDGPU: Fix interaction of tfe and d16

This using the wrong result register, and dropping the result entirely
for v2f16. This would fail to select on the scalar case. I believe it
was also mishandling packed/unpacked subtargets.