Add a notebook with finetuning example

[avidale]

Why ?

In #59, we got asked how SONAR models could be fine-tuned. So I am adding a notebook showing how a SONAR text encoder could be fine-tuned for a classification task.

How ?

• Added a notebook for toxicity classifier training (inspired by MuTox)
• Used an unofficial HF port of the SONAR encoder, to remove the dependency on Fairseq2
• Implemented a custom SonarForSequenceClassification class mimicking similar BERT-based models in the Transformers package.
• Used the HF Trainer for fine-tuning
• Trained two versions: with the encoder frozen and unfrozen (one could also partially unfreeze the top layers)

Test plan

The notebook is self-explanatory and does not need a test.