Fix device synchronization in local_cupy benchmark

[pentschev]

The way the code was written would pass the client's device to each worker and synchronize on that, which is wrong. This was catched due to recent changes to Distributed, where the following was raised:

Task exception was never retrieved
future: <Task finished coro=<Scheduler.broadcast.<locals>.send_message() done, defined at /datasets/pentschev/miniconda3/envs/gdf/lib/python3.7/site-packages/distributed/scheduler.py:4965> exception=Exception("unhashable type: 'cupy.cuda.device.Device'")>
Traceback (most recent call last):
  File "/datasets/pentschev/miniconda3/envs/gdf/lib/python3.7/site-packages/distributed/scheduler.py", line 4969, in send_message
    resp = await send_recv(comm, close=True, serializers=serializers, **msg)
  File "/datasets/pentschev/miniconda3/envs/gdf/lib/python3.7/site-packages/distributed/core.py", line 662, in send_recv
    raise Exception(response["text"])
Exception: unhashable type: 'cupy.cuda.device.Device'

[pentschev]

@charlesbluca make sure you have this patch as well before running anything the local_cupy.py benchmark, or you'll get the errors above as well on latest Distributed.

[pentschev]

Thanks @quasiben for reviewing

[pentschev]

@gpucibot merge