There are quite some changed lines of code, but the overall changes are rather simple:

1. We are now using inside Regression and InferenceProcessor the TextClassificationProcessor.dataset_from_dicts method and only adjust the way labels are converted
2. A bugfix for conversion of python float labels to torch tensors. Previously they were silently converted from float to long, we now have to explicitly check for a correct conversion. This might affect other parts in our code as well. @tholor could you check if we have other float values that are converted to torch tensors?
3. Smaller change in dataset.py for the inferenceprocessing, we now take the whole text as context and not tokens.
4. I moved QAProcessor into NQProcessor, since it was not used elsewhere

partly closes #676