Wide embedding

Changed loss value printed on the screen for regression. From RMSE to MSE. Adjusted documentation. Also changed unit test for callbacks to the new Linear-Embedding implementation, which implied incresing delta for early stopping

#18 Implementation of the Linear model that is the wide component via an Embedding layer. This helps optimize speed and memory usage. Adapted all the submodules accordingly

Precision recall

Changed README, wide_deep.py, docs, and examples according to the new metrics. Only thing left is adjust the notebooks

added link to the docs in the pypi_README file. Note this is a temporal solution. As soon as I manage to include the pngs in the pypi publication this file will dissapear

Build docs

Added a README file so that it looks decent in pypi (I know is not elegant, but I have not managed to properly display png

fixed a bug related to the focal loss. Remove activations before the losses in the fit method in WideDeep. Sigmoid and logSoftmax happen within the loss. Added activations in the train and eval step and in the predict method