Removed single_model_path; added infer_tokenizer to dpr load()

[julian-risch]

Removed single_model_path because we don't use it anymore.
Added parameter infer_tokenizer_classes to load function of DensePassageRetriever so that DPR models with a non-BERT tokenizer can be loaded as required in #1046
The loading can be done as follows after initializing from model hub and saving to disk:

retriever = DensePassageRetriever(
    document_store=InMemoryDocumentStore(),
    query_embedding_model="etalab-ia/dpr-question_encoder-fr_qa-camembert",
    passage_embedding_model="etalab-ia/dpr-ctx_encoder-fr_qa-camembert",
    max_seq_len_query=64,
    max_seq_len_passage=256,
    infer_tokenizer_classes=True
)

save_dir = "../saved_models/dpr"
retriever.save(save_dir=save_dir)

loaded_retriever = DensePassageRetriever.load(load_dir=save_dir, document_store=None, infer_tokenizer_classes=True)

This PR corresponds to FARM PR #765

[tholor]

LGTM

[Timoeller]

The changes are good and solid.

Lets wait with merging until we have the new FARM version in Haystack so we can make use of the infer_tokenizer_class parameter when saving + loading.

[julian-risch]

We have the new FARM version in Haystack, so I am merging now.