AMDGPU: Use attributor to propagate amdgpu-flat-work-group-size

This can merge the acceptable ranges based on the call graph, rather
than the simple application of the attribute. Remove the handling from
the old pass.