Add `gpu_id` option to command line arguments `MLExperiment`

[RobertTLange]

Set cuda visible devices in experiment startup.

[RobertTLange]

Additionally, make it possible to overwrite config file with option. E.g. --lrate 0.001 should set the lrate in train_config.

[RobertTLange]

More general possibility to add device_config to the base configuration file. If it is loaded by the MLExperiment startup, automatically setup the GPU visibility, etc.