Custom data loaders in Base Runner class

[gespyrop]

Hello and thank you for this framework. I have a suggestion that might improve the reusability of the Runner.

After looking at train_nhp_with_features.py, I was wondering whether the TPPRunner (or another Runner subclass) could be used in such custom scenarios.

Using a custom model is already possible since TorchBaseModel.generate_model_from_config searches for TorchBaseModel subclasses named after the given model_id.

However, using custom data loaders is not possible since the Runner directly instantiates a TPPDataLoader which is in turn tightly coupled to the TPPDataset and EventTokenizer classes.

I believe providing a way to inject a custom TPPDataLoader subclass in the Runner would enable it to be used with custom datasets and tokenizers (e.g. TPPDatasetV2 and EventTokenizerV2 from train_nhp_with_features.py).