Support docker_detached for MLPerf inference runs

[arjunsuresh]

So far we are running docker_interactive mode for MLPerf runs. For automated runs, we'll need to use --docker_detached.