This would have the advantage that not every shape would have to copy the reference to the positions, but could e.g. be simply a view on the tensor content itself.
Disadvantage: could be problemativ when being serialized (e.g. similar to map.keySet())