AMDGPU: Fold readlane intrinsics of constants

I'm not 100% sure about this, since I'm worried about IR transforms
that might end up introducing divergence downstream once replaced with
a constant, but I haven't come up with an example yet.

llvm-svn: 363406