Implement Similarity based Answer evaluation

[Timoeller]

Following up on research done in deepset-ai/FARM#757

Target Approach

Implementing a new registered metric using the sentencetransformers model.
Also implement a default metric for FARM with default model. Look for ways to pass comparison model to FARM.

Implementation details

Input to our metric functions, preds and labels, are in number format.
Preds are of type QACandidate, but missing text answer information.
Labels are tuples of start and end indices.

Our evaluator.eval pipeline is not meant for dealing with plaintext predictions or labels because we dismiss our "baskets" with all text information. Also using the tokenizer to detokenize input_ids (token ids) will not work since we do not collect input_ids in the evaluator. Therefore we should start an initial implementation based on Inferencer output and not create a default FARM metric.

[Timoeller]

Fixed by deepset-ai/FARM#803