Bernoulli Naive Bayes

For binary features (word present / absent), Bernoulli naive Bayes models each feature as a coin flip per class, and an absent word counts as evidence too.

Implement BernoulliNB(alpha=1.0) with:

• fit(X, y): class log-priors and per-class feature probabilities $p_{cj} = \frac{N_{cj} + \alpha}{N_c + 2\alpha}$
• predict_log_proba_unnormalized(X): $\log P(c) + \sum_j x_j \log p_{cj} + (1-x_j)\log(1-p_{cj})$ for each class (shape (n, n_classes), classes in sorted order)
• predict(X): the argmax class label