AMDGPU: Update reqd-work-group-size optimization for umin intrinsic

This code was pattern matching the ID computation expression as it
appears in the library. This was a compare and select, but now that
umin is canonical, we were no longer matching. Update to match the
intrinsic instead.