use Uwazi evidences instead of separated training csv file

[Tilana]

use Uwazi evidences instead of separated training csv file to make sure that the training data displayed in Uwazi is consistent with the data that is actually used for training. Also ensure that multiple instances don't share the same training csv file

[Tilana]

requires to pass the evidence sentences of a specific category and value (the same way as in retrain_route()) to predict_one_model() and to predict_route()