tabnet_fit() shall allow to continue a model training with changes hyperparameters.

[cregouby]

This is needed in order to

• allow model fine-tuning ( i.e. manually control a very low learn_rate )
• restart training at a previous checkpoint before model starts to overfit
• allow a future unsupervised training step

This requires splitting model Initialization from model supervised training, while maintaining the capability to save and restore the model at each step.