[meta] Train RTL languages like Arabic and Hebrew

[gregtatum]

RTL languages shouldn't affect training, but doing so will require some work on the Firefox side. This meta bug tracks any work that is needed. We should complete a subset of the easier to segment LTR languages in #524 first as they do not require Firefox changes. These will require a bit more work.

### Tasks
- [ ] [Bug 1876099 - Audit and fix any issues around translating bidirectional language pairs (RTL to LTR, or LTR to RTL)](https://bugzilla.mozilla.org/show_bug.cgi?id=1876099)

There might be some tokenization/segmentation work around Arabic as well.

Native Speakers

If you are a native speaker (L1 language) in any of these languages and want to help out, feel free to leave a comment on this issue or join us in Firefox Translations on matrix. We can always use help with qualitative model evaluation, and questions regarding language.

[BynariStar]

I've found some resources I believe could be useful for training Hebrew models.

At the moment, the best language pair datasets available for Hebrew are the large multilingual ones available on OPUS:

• NLLB
• XLEnt

Here are a few more not on OPUS that might be worth checking as well:

• HebNLI - Manually verified machine translation
• HebWiki QA - Manually verified machine translation
• word2word - A simple word-to-word pair dataset
• Hebrew WordNet (Archived) - The website seems to be down. I will try to see if they have a backup.

Some extra resources:

• HebSpacy - An NER model used by Azure AI Language for TA4H in Hebrew. Made in collaboration between Microsoft, the Israeli Ministry of Health, an Israeli HMO, and others.
• The ONLP Lab - An NLP research lab at the Bar Ilan University in Israel. They have a lot of Hebrew NLP resources and models.
• Hebrew NLP Resources - A list of Hebrew NLP resources compiled by NNLP, an Israeli government Hebrew NLP advancement project.