Allow datasets to be loaded from the Hugging Face Hub

[lewtun]

Hello, currently all the logic to load datasets is hard-coded in pipelinerl/domains/math/load_datasets.py. It would be nice if one could specify the train and test datasets to point directly to ones on the Hugging Face Hub, like this:

train_dataset_names:
    - {ORG}/{TRAIN_DATASET_NAME}
test_dataset_names:
    - {ORG}/{TEST_DATASET_NAME}

As long as the datasets are preprocessed in the expected format, we could skip the hard-coded logic in load_datasets.py altogether and make the framework a lot more flexible / user friendly :)

[lewtun]

Given that Hub datasets have subsets/configs and splits, a better design could be:

train_dataset_names:
  - hub_id: {ORG}/{TRAIN_DATASET_NAME}
    config: foo
    split: train
test_dataset_names:
  - hub_id: {ORG}/{TEST_DATASET_NAME}
    config: bar
    split: test