FEAT Add SurvTRACE

[Vincent-Maladiere]

This PR aims at refactoring and packaging the SurvTRACE model. This effort focuses on:

• Generalizing the preprocessing and feature engineering of SurvTRACE so that we can run it on any dataset (instead of the three provided).
• Decoupling the preprocessing from the model, which relies on a STConfig global dictionary. We want to remove this global config and use hyper-parameters the scikit-learn way.
• Simplifying the codebase and removing some of the boilerplate by using skorch. Also, skorch will enable pipelining and cross-validation operations.

cc @ogrisel

[Vincent-Maladiere]

This a "working WIP", upon running fit I get similar performances for training and validation errors as in the original SurvTRACE version.

Next steps:

• [ ] update the tests and complete them
• [ ] get the final metrics and compare results with the original SurvTRACE version.

Usage (with the seer dataset named "seer_cancer_cardio_raw_data.txt"):

from hazardous.data._seer import (
    load_seer,
    CATEGORICAL_FEATURES,
    NUMERICAL_FEATURES,
)
from hazardous.survtrace._model import SurvTRACE

X, y = load_seer("hazardous/data/seer_cancer_cardio_raw_data.txt")
print(X.shape, y.shape)  # (476746, 28), (476746, 2)

model = SurvTRACE(
    numerical_features=NUMERICAL_FEATURES,
    categorical_features=CATEGORICAL_FEATURES,
)
model.fit(X, y)

cc @ogrisel, I think this is reviewable :)

[Vincent-Maladiere]

This PR is being split:

• the SEER part is being developed at https://github.com/soda-inria/hazardous/pull/24
• the SurvTRACE part is being developed in the branch icml-2024

[ogrisel]

Let's close this and later open a new PR dedicated to the model then.