Rewrite guides for fine-tuning with Datasets

[stevhliu]

What does this PR do?

This PR updates the current documentation for Fine-tuning with custom datasets. It removes the custom code for loading a dataset in favor of using the Datasets library for loading and preprocessing a dataset. The new guide also introduces the Keras method for compiling and fitting a model instead of using TFTrainingArguments and TFTrainer.

⚠️ The TensorFlow example for Question Answering still needs a little work. It returns ValueError: No gradients provided for any variable when calling model.fit. @Rocketknight1 has provided a temporary solution where we set dummy_labels=True in tf_to_dataset.

This notebook contains all the code examples shown in the guide.

To do:

• Update the TensorFlow code example for question answering once we have a better solution.

[sgugger]

Thanks a lot for working on this! I have a general comment on the code samples. The >>> are usually used to go with the doctest package, but it also requires to have ... on all the lines not having >>>, which is not the case here.

We should either pick if we want this guide to be enabled for the doctests (in which case we should add the ...) or not (in which case the >>> just hurt readability and should be removed).

[LysandreJik]

Nice! The content looks great to me. I think there's a good opportunity to do something a bout doctests here, which would tremendously help in maintaining this guide.

We have setup doctests for a few files but we'll need to enable them for more files - @stevhliu let us know if you'd like us to walk you through how these work and how these are setup so that your work may be tested, which should greatly reduce the maintenance cost down the road.

[LysandreJik]

Thank you, @stevhliu!