T5 support

[davidkaczer]

What does this PR do?

This PR adds support for pretrained LongT5 encoder-decoder models from HuggingFace.

General Changes

• Add HuggingFacePretrainedEncoderDecoderModel class for loading LongT5
• Add span masking collator ported from official Google implementation
• Add test for span masking collator
• Add example training configs for LongT5
• Add logic to distinguish between decoder-only and encoder-decoder models when passing inputs

Breaking Changes

• None intended

Checklist before submitting final PR

• [x] My PR is minimal and addresses one issue in isolation
• [x] I have merged the latest version of the target branch into this feature branch
• [x] I have reviewed my own code w.r.t. correct implementation, missing type hints, proper documentation, etc.
• [x] I have run a sample config for model training
• [ ] I have checked that all tests run through (python tests/tests.py) (apparently some tests were failing already in upstream? also didn't test multi-GPU)
• [ ] I have updated the internal changelog (CHANGELOG_DEV.md)