Fix some minor type and style issues. Adjusted the ZILN Loss so it works with former versions of pytorch and rename quantile regression multilabel label as qregression

Fixed #60 and #61. Placing dropout in the right position and applying FullEmbeddingDropout only in training

Against all good practices, this is a massive commit that adds an entire new module with Bayesian models (that are still not functional). Making them functional requires some work. Also the entire models module has been restrucured in preparation for better days to come

need consultation, I am not sure if I got it correctly

First commit towards v2. Re-organized the models module and added a few new functionalities for the models in there

added embedding rules, MonteCarlo(uncertainty) prediction and removed running of the tests on draf requests

Removed condition from trainer on duplicated callbacks. Tab2Vec optionally returns pandas dataframes. Added unit tests for new functionalities and all notebooks follow now black coding style. All this relates to PR #55

code cleaned + added workaround for ¨ImportError: libGL.so.1: cannot open shared object file: No such file or directory ``` apt-get update apt-get install ffmpeg libsm6 libxext6  -y ```¨

* added RayTuneReporter callback
* adjusted history so it reports metrics per class in case of multiple values per metric. eg. multiple classes

* added notebook and 1 report for metrics

Added docs. Re-design Additive Atttention accoring to their code. Add a few details on the docs and fix a typo

 Please enter the commit message for your changes. Lines starting

Added a proper implementation of transformer models when needed and re-organised the transformers module for clarity