BERT Tests

[dilyabareeva]

This is part of an effort to adopt quanda as a language modality. Ideally, we want our explainers to support the HuggingFace AutoModelForSequenceClassification model instances (see also #8). I don't know if any of our explainer wrapper classes currently support it.

Step for this issue:

• Add a model, such as this, in combination with its train dataset, as a test suite -> Can we find a "mini" version of the model? Can we only add a subset of the train set? The tests should be as quick as possible in the end, and take as little memory as possible.
• If needed, adjust the TRAK wrapper class to be able to process this model and the dataset. See https://github.com/MadryLab/trak/blob/main/examples/qnli.py.
• In the TRAK example, they create a wrapper for an AutoModelForSequenceClassification. Can we incorporate this wrapper into our code base, so that a TRAK explainer can be initialized with an AutoModelForSequenceClassification model? Please figure out the best solution here.
• Build TRAK wrapper tests for the newly added language test suit.
• Adjust the base Explainer accordingly.