Add backend arg

[Delaunay]

Intel Gaudi & GPU Max come with their own dist backend (hccl, ccl respectively). This patch enable those GPUs to be used in parallel to speed up training

[rwightman]

@Delaunay thanks, should be there yeah, maybe default to None though so a device based default can be used if not explicitly set?

[Delaunay]

I set the default to none