Add SentenceTransformersRanker with pre-trained Cross-Encoder

[julian-risch]

In contrast to FARMRanker, SentenceTransformerRanker uses the logit as similarity score and not the classifier's probability of label "1"
see example here: https://www.sbert.net/docs/pretrained-models/ce-msmarco.html#usage-with-transformer

I tested with a subset of the nq_dev dataset. Here are the results of a pipeline with ElasticsearchRetriever and SentenceTransformerRanker with "cross-encoder/ms-marco-MiniLM-L-12-v2" as model:

EvalRetriever
-----------------
has_answer recall@2: 0.7200 (18/25)
no_answer recall@2:  1.00 (25/25) (no_answer samples are always treated as correctly retrieved)
has_answer mean_reciprocal_rank@2: 0.6200
no_answer mean_reciprocal_rank@2:  1.0000 (no_answer samples are always treated as correctly retrieved at rank 1)
recall@2: 0.8600 (43 / 50)
mean_reciprocal_rank@2: 0.8100

Retriever (Speed)
---------------
No indexing performed via Retriever.run()
Queries Performed: 50
Query time: 0.3390099899999086s
0.0067801997999981725 seconds per query

EvalRanker
-----------------
has_answer recall@2: 0.7600 (19/25)
no_answer recall@2:  1.00 (25/25) (no_answer samples are always treated as correctly retrieved)
has_answer mean_reciprocal_rank@2: 0.6600
no_answer mean_reciprocal_rank@2:  1.0000 (no_answer samples are always treated as correctly retrieved at rank 1)
recall@2: 0.8800 (44 / 50)
mean_reciprocal_rank@2: 0.8300

Ranker (Speed)
---------------
Queries Performed: 50
Query time: 161.3018365920002s
3.226036731840004 seconds per query

Limitations: documentation on the website has not been updated. It might be unclear/confusing for users at the moment whether to use FARMRanker or SentenceTransformerRanker.

closes #1129

[tholor]

I think the separation makes sense (at least for now). We might combine them later on and rather add an arg similarity_type to differentiate between the two different approaches.

Please add some more documentation (see comments) and a basic test case for both rankers that ensures the expected scores / sorting of some dummy docs (FARM + sentencetransformers).

[julian-risch]

@tholor Thank you for your feedback! I made the requested changes.

[tholor]

Looking good! Thx for the changes.

I added a minor sentence to the docstring (we should keep in mind that some user might not know what Re-ranking is and explain at least very briefly the "value / use case".

Also added the import to the init so that we can just import from haystack.ranker import SentenceTransformersRanker similar to our other building blocks.