feat(naive_bayes): add CategoricalNB classifier

[uttam12331]

Closes #1399.

Adds an online CategoricalNB naive Bayes classifier for categorical features (e.g. {"weather": "sunny", "humidity": "high"}), mirroring scikit-learn's CategoricalNB but learning incrementally. It fills the gap alongside the existing MultinomialNB/BernoulliNB/ComplementNB/GaussianNB.

Model

For each feature f, category v and class c:

P(x_f = v | c) = (count[f][(c, v)] + alpha) / (class_count[c] + alpha * n_categories(f))

with the empirical class prior P(c) = class_count[c] / N. New categories seen after the first observations are handled gracefully (online setting).

Implementation

• learn_one / learn_many (mini-batch via MiniBatchClassifier), joint_log_likelihood / joint_log_likelihood_many, plus p_class / p_feature_given_class helpers — consistent with the other NB classes.
• Registered in naive_bayes/__init__.py.

Verification

• Matches scikit-learn's CategoricalNB to machine precision (max predict_proba difference ~4e-16 across alphas) — covered by a new test_categorical_vs_sklearn (parametrized over alpha).
• learn_many produces an identical model to repeated learn_one (test_categorical_learn_many_vs_learn_one).
• Unseen categories at predict time don't raise and keep probabilities normalized (test_categorical_handles_unseen_feature_value).
• A runnable doctest example in the class docstring.
• Full river/naive_bayes suite passes (50 tests), the generic estimator checks pass (test_estimators.py -k CategoricalNB, 53 checks), and ruff check / ruff format are clean.

[MaxHalford]

Nice addition!

[MaxHalford]

Could you support Narwhals? We're in the process of moving all mini-batch methods to Narwhals instead of pandas

[MaxHalford]

We do not want mini-batch methods to be for loops over the inputs. Mini-batch methods must use vectorization, else they're not bringing anything to the table.