Reuse user process group.

[wujingyue]

A user process group has to be created before nvFuser to even initialize a device mesh. Currently, nvFuser's communicator creates its own process groups listening to a different port. This wastes resources and probably has triggered some conflicts that lead to NVFUSER_DISABLE=multidevice.

cc @syed-ahmed: I'm pretty sure I asked you about this and then forgot. Did you recommend a way to register the user process group that's accessible to the C++ side of nvFuser?

[syed-ahmed]

You can register the user process group from python like this:

https://github.com/pytorch/pytorch/blob/59b7e52ad8f6146b4364515a7f3e54d6f3edd6da/test/distributed/test_c10d_functional_native.py#L86-L92

You can then resolve the process group on the C++ side using the name like this:

https://github.com/pytorch/pytorch/blob/59b7e52ad8f6146b4364515a7f3e54d6f3edd6da/torch/csrc/distributed/c10d/Functional.cpp#L43-L44