support for running containers on nodes with GPU

[anaderi]

to run code inside container that runs something on nvidia GPU, one has to take care of certain options for docker that tell it to bind device and drivers inside container (of course image it run should be created accordingly -- TODO: describe on wiki).
there is a helper project that takes care of everything: https://github.com/NVIDIA/nvidia-docker, it creates a special service (listens to http://localhost:3476 and tells additional arguments to docker: https://github.com/NVIDIA/nvidia-docker/wiki/nvidia-docker-plugin, i.e. http://localhost:3476/docker/cli/json returns json with all additional args that would enable running GPU-specific stuff inside container, i.e. nvidia-smi)
in particular there are devices arguments that docker-py of version 2.0+ can understand.
So the task would be to

• upgrade docker-py to 2.2.1
• if running on single docker (not swarm) - check nvidia-docker-plugin availability, and if it is there, get the arguments and pass it to docker.client.run
• if running on docker swarm, there is a blocker: NVIDIA/nvidia-docker#141