feat(l2g): better l2g training, evaluation, and integration

[ireneisdoomed]

✨ Context

L2G training and prediction rewrite to use Scikit instead of pyspark.ml.

The main points that made me do this were:

• Evaluation of the model was really slow. Even though I was persisting the L2G feature matrix and model, any time that we interacted with the model, this was triggering a retraining. This led to runtimes of ~2h if we wanted to run the pipeline end to end.
• I don't think using Spark is necessary for training. We have a simple model. Testing the approach to scikit made training and predicting possible in seconds.
• Using scikit brings many niceties: bigger support from community, more control, native integration with W&B (panel is more informative now), the model is easier to share, and we can upload it to the Hugging Face Hub.

Training end to end is now 18mins - job
Predicting is also 18 mins - job

This basically reassures me that annotating features is practically the entirety of the pipeline, which is consistent with my tests.

🛠 What does this PR implement

I've deprecated almost its entirety, so it's easier if i describe what this inclu
However, the L2G workflow is very similar to the previous one. This is just a change in the implementation that optimises the process.

• LocusToGeneStep.run_train creates a L2GModel based on a GradientBoostingClassifier. The model is trained, exported to Hugging Face and saved locally.
• LocusToGeneStep.run_predict downloads the model from the hub by default, and extracts scores.
• The W&B and Hugging Face secrets are now accessed from the Secret Manager. So this part is now fully integrated in the pipeline.
  • Added the copy_to_gcs util to move the L2G model local file into a GCS destination.
• The only particularity in LocusToGeneTrainer is the addition of hyperparameter_tuning. This function uses Weights&Biases Sweeps, to sweep over all parameters provided in a grid, train a model, evaluate the results and upload them in a group. For example:
  
  (interesting that the metrics are the same) This is not part of the pipeline, only useful in development stages.
• L2GFeatureMatrix have 2 new attributes fixed_cols, to complement the columns that are not features, and mode. Both are connected, on train mode the fixed columns are studyLocusId, geneId and the target column (goldStandardSet); on predict mode, we don't have the latter one.
• LocusToGeneModel gain 3 attributes: hyperparameters, training_data (to store the input feature matrix and log it to W&B), label_encoder. Features_list is no longer necessary.

🙈 Missing

• Bug in the logging of the feature matrix that don't show the feature labels
• Cross validation to have a more accurate evaluation of the model.
• Testing for the Training functions
• When the model is public, we could upload the predictions dataset to the HF Hub to link a dataset with a model.

🚦 Before submitting

• Do these changes cover one single feature (one change at a time)?
• Did you read the contributor guideline?
• Did you make sure to update the documentation with your changes?
• Did you make sure there is no commented out code in this PR?
• Did you follow conventional commits standards in PR title and commit messages?
• Did you make sure the branch is up-to-date with the dev branch?
• Did you write any new necessary tests?
• Did you make sure the changes pass local tests (make test)?
• Did you make sure the changes pass pre-commit rules (e.g poetry run pre-commit run --all-files)?