Add Roberta, XLNet and redesign Tokenizer

[tholor]

We have more models joining the FARM: XLNet and Roberta 🎉
This PR includes fundamental changes in FARM to make it easier to add other language models.

Support for tokenizers from the transformers repo.

Pros:

• It's quite easy to add a tokenizer for any of the models implemented in transformers.
• We rather support the development there than building something in parallel
• The additional metadata during tokenization (offsets, start_of_word) is still created via tokenize_with_metadata
• We can use encode_plus to add model specific special tokens (CLS, SEP ...)

Cons:

• We had to deprecate our attribute "never_split_chars" that allowed to adjust the BasicTokenizer of BERT.
• Custom vocab is now realized by increasing vocab_size instead of replacing unused tokens

Generic loading of tokenizers:

BertTokenizer.from_pretrained("bert-base-cased") -> Tokenizer.load("bert-base-cased")

Generic loading of language models:

Bert.load("bert-base-cased") -> LanguageModel.load("bert-base-cased")

Using other models is as easy as switching the name: LanguageModel.load("roberta-base")

Downstream tasks

So far, the new models are available for text classification, multilabel text classification, regression and NER.
For QA and LM finetuning the processors have still to be updated.

[Timoeller]

Looking good. I like how you adjust the model inputs in the forward pass of each LM model type.

Lets delay the discussion about the whitespace tokenization + normal tokenization afterwards and continue. It seems it doesnt change in the current PR.

Lets go forward with this, very amazing to have other models in here.

[brandenchan]

Looks good