Decrease number of required fields in TrainingArgs and TorchrunArgs

Currently we have two main objects that are used when launching training. These require a set amount of options to be specified. But the training library should infer a number of these when no values are provided so that it's easier to consume.
