Clarify the intended semantics of attach vs offset in version >= 5

As can be seen in https://gitlab.gnome.org/GNOME/mutter/-/issues/2622, using wl_surface_offset did not actually fix the problem with DND hotspots in GTK. That is because the implementation in both mutter and weston checks that dx/dy are 0 if version >= 5, but then go ahead and overwrite the existing offsets with that 0 anyway.

The intention of the protocol addition clearly was not to introduce an ordering dependency between attach and offset. It would be good to clarify the wording in the protocol to make it clear that attach ignores dx/dy if version >= 5 (still requiring them to be 0).