[Parent] Docketbert pretraining

[nadahlberg]

Pretrain a domain-adapted modernbert for docket entry description text.

Run initial experiments on a small set of ~6M entries to test different strategies including:

• Vanilla masked language modeling with modernbert models
• Finetuning small models from scratch
• Finetuning models with a distillation loss + MLM
• Finetuning "sliced" variants of modernbert models

Then run a larger training run with ~50M entries using the best strategies.