Added option to specify list of module names instead of single module name

[lucaeyring]

This PR adds the functionality to save multiple intermediate features in a single forward pass.

This is implemented through a new argument module_names. Now either module_name or module_names need to be supplied. This is done for backwards compatibility (everything should still work with just module_name).

One change needed to be made, which is that features is now a Dict from module_name to corresponding features_module, and the filename, which the features are saved under, now includes the module_name. Need to double check whether this breaks anything w.r.t. backward compatibility.

[lciernik]

I left some comments on where things could be done slightly differently.

Some questions remain:

• Where is the backward compatibility logic when module_name is given? It appears to me that _extract_batch will, in any case, return a dictionary of module names and features.
• We can apply alignment of representations. See align method in torch.py. Does this still work with multiple module_names?
• What happens if the dataset is huge?

[lciernik]

Still have some comments.

[LukasMut]

This fixes issue #89