Running NVidia dockers

[manthey]

Currently, if a docker task has the appropriate labels, we try to run the docker task with the runtime='nvidia' parameter. As of Docker 19.03, there is docker built-in support for the --gpus flag (more generally, for a device request). As soon as docker-py supports this (see issue https://github.com/docker/docker-py/issues/2395 and PRs https://github.com/docker/docker-py/pull/2419, https://github.com/docker/docker-py/pull/2471), we should switch to using the appropriate method. Until then, there are extra steps to use nvidia dockers.