(suggestion) tensor data pointer GPUs

[janbrandejs]

Should the pointer to tensor data be a struct (instead of a bare pointer) containing the pointer and then an information about the target device?
For cases where there is no unified address space of CPUs and GPUs.
(comment from Anthony Scemama)