Transformers to 2.4/2.5

[sleepinyourhat]

• Revisit major new Transformers features—consider adopting their methods for checkpointing etc. if there's a clear advantage.
• In any case, update the required package version, and make sure it runs with both current and new pre-trained models.

[zphang]

To add urgency to this, XLM-RoBERTa is basically broken in v2.3.0. It cannot be loaded in jiant at all because of an issue in the tokenizer vocabulary (Core dump due to a bad vocabulary lookup, because of an issue with vocab size; we exhaustively look up the whole vocabulary when setting up our models).

See: huggingface/transformers#3014 (review) for more details

We will likely want to update to at least v2.5.1 (v2.5.0 seems to have its own set of issues).

[pyeres]

Addressed by @zphang's PR (#1059) which updated dependency to Transformers version 2.6.0.