Pipeline training dataset refactoring

[alaeddine-13]

Currently, our training pipeline supports only 1 dataset that conform to the alpaca format by using the dataset name. We need to ensure that:

• [ ] All datasets are remote and pulled from huggingface at run time
• [ ] Add flexibility to the training pipeline so that we can specify datasets with different format using defined preprocessors
• [ ] Allow specifiying and concatenating multiple datasets