This allows us to shrink the fellow_cache_obj allocation for the
fellow_obj_get (from disk) case from 8KB to 512 bytes.

The root cause for the massively oversized fco allocation was that
the nseg_guess heuristic could not take into account the wsl (size of
the actual object), so it had to assume that all of the disk object's
size was taken up for segments in the disk object embedded seglist.