Support for multicast/unicast?

This came up in a discussion whether one could support multicast/unicast in DLPack.

I don't think this is used currently by things like CuPy/torch but other users seem to have the need to share this information.
As I am not very familiar with multicast/unicast, I think a multicast array would actually be it's own device, while a unicast could be a normal CUDA allocation, but unicast is still good information.
If unicast here is sufficient as _optional_ metadata, then it could be a flag. Although I guess it might get us into a place where we may want device specific flags.

If there is a use-case for adding devices, that seems just very cheap to me (a simple enum entry after all), flags also are, although if we want to do device-specific flags/metadata we may want to extend the struct (or reserve some of the flag space).

No files or tests are named. Start by reviewing the current device enum, tensor struct, and flag definitions, then clarify whether multicast requires a new device and whether unicast belongs in optional metadata; done when the representation and compatibility scope are agreed.

Written by the indexing model from the issue text.