Does not include a handler's tensor representation (and so no copy-on etc.), and almost all of the hooks are missing.

My medium-term goal is to get the parallel device working with function replay so TPU collectives work inside functions. That will also get us a replication primitive for use with the eager/graph agnostic C API, and I'll plan to call it from the existing custom device to start.

PiperOrigin-RevId: 340253840
Change-Id: Ic9a5acca7bf42ceb9cb54aca635a9861daca3b38