Hopsparser

Fixed

- Update to pydantic 2 to fix breaking changes and pin it.

Added

- HOPS now provides custom spaCy component to use in spaCy pipelines.
- Options for using weighted multitask losses, including the adaptative strategy used in Candito
([2022](https://aclanthology.org/2022.findings-acl.190)).
- HOPS will learn and predict token-level labels encoded in the MISC column (as `key=value`) if you
give it the name of the key in the `extra_annotations` part of the config. See the example
mDeBERTa-polyglot
[config](https://github.com/hopsparser/hopsparser/blob/main/examples/mdeberta-polyglot.yaml) for
an example of such a config

Added

- `hopsparser evaluate` now accepts an optional output argument, allowing to write directly to a
file if needed.
- [A new script](test_models.py) to help catch performances regressions on released models.

Changed

- We now accept partially annotated CoNLL-U files as input for training: any learnable cell (UPOS,
HEAD, DEPREL) for which the value is `_` will not contribute to the loss.

The performances of the contemporary models in this release are improved, most notably for models
not using BERT.

Added

- The `scripts/zenodo_upload.py` script, a helper for uploading files to a Zenodo deposit.

Changed

- The CharRNN lexer now represent words with last hidden (instead of cell) state of the LSTM and do
not run on padding anymore.
- Minimal Pytorch version is now `1.9.0`
- Minimal Transformers version is now `4.19.0`
- Use `torch.inference_mode` instead of `toch.no_grad` over all the parser methods.
- BERT lexer batches no longer have an obsolete, always zero `word_indices` attribute
- `DependencyDataset` does not have lexicon attributes (`ito(lab|tag` and their inverse) since we
don't need these anymore.
- The `train_model` script now skips incomplete runs with a warning.
- The `train_model` script has nicer logging, including progress bars to help keep track of the
experiments.

Fixed

- The first word in the word embeddings lexer vocabulary is not used as padding anymore and has a
real embedding.
- BERT embeddings are now correctly computed with an attention mask to ignore padding.
- The root token embedding coming from BERT lexers is now an average of non-padding words'
embeddings
- FastText embeddings are now computed by averaging over non-padding subwords' embeddings.
- In server mode, models are now correctly in eval mode and processing is done
in `torch.inference_mode`.

Fixed

- Model cross-device loading (e.g. loading on CPU a model trained on GPU) works now ([65](https://github.com/hopsparser/hopsparser/issues/65))

Changed

- Remove the dependency on `click_pathlib` ([63](https://github.com/hopsparser/hopsparser/pull/63))

Fixed

- Compatibility with setuptools 61 parsing of PEP 621 specs