[meta] Translate paragraphs instead of sentences

[eu9ene]

Based on WMT24, sentence-level translation is going away. There's now more document-level training data available (for example HPLT), and WMT24 used document-level datasets for evaluation.

See Findings of WMT 2024 Shared task

In a shift towards document-level evaluation, we
no longer provide source texts segmented into indi-
vidual sentences. Instead, we keep all paragraphs
intact and evaluated together.

This would require:

• adapting document level datasets to leave some paragraphs to train on instead of splitting to sentences
• fix cleaning procedures
• find evaluation datasets
• implement inference support