Incorporate cross-device-sharing into the memory model in `Tnode`

[lukstafi]

In particular, this will make it explicitly controllable and more visible. Currently, the classification is implicitly implemented in the CUDA backend.