Add convenient Crepe.represent_sequences method

[willdumm]

This addresses #115, adding Crepe.represent_sequences, and a number of supporting methods on D*SM model classes.

It also eliminates the option of providing non-paired sequences to D*SM model methods that take a string (and therefore also all Crepe methods that call them). These methods now require amino acid sequences to be provided in (heavy_chain, light_chain) tuples, where a missing chain sequence can be represented by the empty string.

The represent_sequences function returns a tensor for each heavy-light pair provided to it, while Crepe.__call__ returns a pair of tensors (one for heavy, one for light chain) for each heavy-light pair provided to it. This seems to me the correct choice, but there could be justification for splitting the embedding tensors returned by represent_sequences on heavy/light boundaries.

[willdumm]

This function had to be moved to avoid circular dependencies.

[matsen]

Great!

Q: If we can't supply single chains, does it make sense to require paired tuples? Ablang take lists of lists and I can get those with

one_pair = [train_df.iloc[0][['heavy', 'light']].tolist()]

Of course, I can manage with

def lists_to_tuples(list_of_lists):
    return [tuple(lst) for lst in list_of_lists]

rep = crepe.represent_sequences(lists_to_tuples(one_pair))

No big deal but I thought I'd ask.

Also, it appears that the return type of represent_sequences is a tuple. Is that on purpose?

[willdumm]

Thanks for noticing these things! I changed the check on sequence inputs to allow any non-str type of length two. Also, I changed the return type to list.

[matsen]

Works like a charm. Merge it!