Automatically determine an appropriate number of PyTorch workers

[kamilzyla]

The option --pytorch_num_workers must be set to 0 on Windows; otherwise the classifier script fails in a cryptic way (see #90). A better solution would be to automatically determine a proper number of workers based on the system (keep the option so the user can overwrite this setting).

In addition it would be good to research what setting is actually appropriate and document the choice. Does having more workers make the inference faster? Does it come with increased memory usage (it might lead to crashes, see #74) or other drawbacks?