integrations: Load `model_file` for resuming

[daavoo]

In order to properly resume training with dvc checkpoints, the user needs to load the existing model_file at the beginning of training.

Given that DVCLive integrations already take care of saving the model_file I think it makes sense to also include some logic to load the model_file, if it already exists, on the callback instantiation or on_train_begin.

This would simplify the usage of dvc checkpoints for resuming training.

• Catalyst
• Fastai
• HuggingFace
• keras Keras load model #174
• LightGBM
• MMCV
• PyTorch
• PyTorch Lightning: Support saving model to model_file #170
• TensorFlow
• XGBoost

[daavoo]

It looks that for some ML Frameworks which have some "resuming" argument this should be a Documentation update

[daavoo]

Lowering the priority as this depends on deciding and documenting the recomended workflow for DVC checkpoints