Add "no answer" aggregation to Transformersreader

[Timoeller]

Linked to #257 and #258

Right now the Transformersreader just gives back "no answers" for each document it predicts on.

We need to aggregate those "no answers" and their scores, similar to what the FARMReader is doing.

The Transformersreader pipeline does not seem to have an option to

1. modify the "no answer" bias
2. always return the "no answer" (no matter how unlikely it is)

Because of these limitations we need workarounds.

[Timoeller]

I added the aggregation functionality.
The aggregation only works correctly when "no answer" is most likely across all documents. In all other cases the "no answer" is not correctly sorted into our list of predictions. This happens because transformers not always returns the "no answer"