Describe the workflow you want to enable

There is an ongoing discussion on #22574 about introducing a new estimator named ColumnwiseNB, which aims to handle different types of features by applying different Naive Bayes models column-wise. This approach is promising for datasets that contain a mix of categorical, binary, and continuous variables, each of which might require a different probabilistic approach for effective classification.

from sklearn.naive_bayes import BernoulliNB, GaussianNB, CategoricalNB

clf = ColumnwiseNB(nb_estimators=[('gnb1', GaussianNB(), [0, 1]),
                                  ('bnb2', BernoulliNB(), [2]),
                                  ('cnb1', CategoricalNB(), [3, 4])])
clf.fit(X_train, y_train)
clf.predict(X_test)

Describe your proposed solution

While scikit-learn is considering the ColumnwiseNB as a potential addition, I've developed a similar feature for a while called GeneralNB in the wnb Python package. This class also supports different distributions for each feature, providing flexibility in handling a variety of data types within a Naive Bayes framework. I would like to introduce the community to this already-implemented solution to gather feedback, comments, and suggestions. Understanding whether GeneralNB could serve as a good alternative or complementary solution to ColumnwiseNB could be beneficial for both scikit-learn developers and users looking for advanced Naive Bayes functionalities.

from wnb import GeneralNB, Distribution as D

gnb = GeneralNB(
    distributions=[D.NORMAL, D.NORMAL, D.BERNOULLI, D.CATEGORICAL, D.CATEGORICAL])
gnb.fit(X_train, y_train)
gnb.predict(X_test)

This solution fully adheres to scikit-learn's API and supports the following continuous and discrete distributions at the moment of writing this issue:

• Normal
• Lognormal
• Exponential
• Uniform
• Pareto
• Gamma
• Beta
• Chi-squared
• T
• Rayleigh
• Bernoulli
• Categorical
• Geometric
• Poisson

I encourage community feedback on this implementation and am open to collaborating to integrate similar functionality into scikit-learn if deemed beneficial.

Describe alternatives you've considered, if relevant

No response

Additional context

No response