Scale dot product into probabilities

[brandenchan]

This PR makes improvements and fixes around the similarity functions used in embedding based retrieval.

A warning will be logged in cases where there is a non-recommended Dense Retrieval Model / similarity function pairing.

Scaling of dot product scores into probabilities is now also supported.

This addresses #661

TODO:

• Raise warning with bad pairing (ES)
• Scale dot product properly
• Add similarity as attribute of DocumentStore mother class
• Implement both similarity fns in InMemoryDocumentStore
• Implement dot product in FAISSDocumentStore
• Implement neither in SQLDocumentStore (But raise warning?)
• Give each document store its own _create_document_field_map() method
• Compare numbers across doc stores (To be addressed in Align similarity functions across DocumentStores #672)

[tholor]

Looking good! Just one comment on the docs / logging + mypy is still failing ..

[brandenchan]

Implemented support for each document store and similarity function but the probabilities being returned are inconsistent. Will be addressed by #672