Bio-transformers

Features:
- Add msa-transformers for methods:
- compute_logits
- compute_embeddings
- compute_probabilities
- compute_accuracy

Fixed:
- Remove torch DataParallel wrapper.

Features:
- Add ray worker for multi-gpus inference

Removed:
- Remove torch DataParallel wrapper.

Note on the release

Features:
- Add BIO_LOG_LEVEL environnement variable to control logging message (logger)
- Check if every unique amino acids in sequences are in tokens_list (compute_probabilities)

Fixed:
- Add shuffling in batch_sampler (lightning_utils)
- Fix tokens argument for dataloader (lightning_utils)
- Fix rtd CI to separates docs and package environment.

Changed:
- Modified the signature of some functions to improve clarity (tansformers_wrappers)
- Update `train_masked` method to `finetune` (tansformers_wrappers)
- `compute_embeddings` with option `full` return a list of embeddingsn, no matter the size (tansformers_wrappers)

Removed:
- Remove the tokens_list argument when not necessary and tried to make its usage clearer (tansformers_wrappers)
- Remove functions (tansformers_wrappers):
- _filter_and_pool_embeddings
- _split_logits
- _slabels_remaping
- _filter_logits
- _filter_loglikelihood
- _compute_accuracy
- _compute_calibration

Fixed:
- Batch_sampler issue

Note on the release

Features:
- Merge ESM/protbert for finetuning model with pytorch-lightning
- Possibility to restore a training session.

Fixed:
- Fix conflicts when saving model with DDP
- Fix loading checkpoint created by pytorch-lightning

Note on the release

Features:
- Add fasta files support for each compute function.
- Add train_masked function to finetune model on custom dataset. (Only ESM for the moment, protbert is coming.)

Docs:
- Update documentation to add tutorial on training.

Changed:
- GPU is used by default if found, even if not specified.