Refactor pipeline for better generalizability & Add TransformersReader

[tholor]

Add an option to use QA models from huggingface's Transformers as a Reader.

Advantages:

• users can leverage quite easily other community models (from here: https://huggingface.co/models)
• simpler workflow for transformer users to use their own models (no need to convert to FARM model)

Hand in hand with this addition, it became clear that we want the pipeline to be more generalizable:

• let the retriever return documents and not only indexes+scores
• don't require the retriever to have a df in memory with the docs (needed for scaling, e.g. with elasticsearch backend)
• make input to reader.predict more explicit
• simplify the format of the dict returned by reader.predict