DeezyMatch training requires training dataset to exist even if training is skipped

[mcollardanuy]

Currently, if a DeezyMatch model already exists, the Ranker (when in "deezymatch" mode) still goes through the code to generate a training set for DeezyMatch. It shouldn't, because we already have a DeezyMatch model. This may cause an error if we use DeezyMatch with an already trained DeezyMatch model in an inference setting only (i.e. in a location where we haven't run the training), because the code looks for the resources to generate a training set, even if we don't provide them because we already provided the trained model.