Investigate using an LLM as a teacher

[eu9ene]

We should do a proof of concept for knowledge distillation from an LLM to our standard student model. The main benefit of this if it works is we won't need to deal with parallel data cleaning and training teacher models. All this can be quite challenging, especially for lower-resource languages.

This would require:

• Estimate costs for different LLMs and APIs we can use
• Run quality evaluation for those to see which model would provide the best cost/quality tradeoff
• Choose a mix of monolingual data to translate
• Run translation with an LLM
• Train a regular student model on this data
• Try different LLMs and corpus of different sizes (for example, 10M and 50M sentences)

Folks from WMT also suggested we can try pre-training the student on parallel OPUS corpus as is and then finetune on a smaller but high-quality LLM-produced corpus to make it more cost efficient.

[eu9ene]

Related to #767