Fix HNOCA kernel to be the squared Jarrad kernel

[Marius1311]

Problem

The HNOCA kernel is currently computed as:

kernel_matrix.data /= 2 * n_neighbors - kernel_matrix.data
kernel_matrix.data = kernel_matrix.data**2

This means the base kernel before squaring uses 2 * n_neighbors as the denominator, while the Jarrad (Jaccard-based) kernel uses 4 * n_neighbors:

kernel_matrix.data /= 4 * n_neighbors - kernel_matrix.data  # Jarrad / Jaccard

Fix

To be faithful to the original HNOCA implementation, the HNOCA kernel should be the Jarrad kernel squared, i.e.:

HNOCA(x, y) = Jarrad(x, y)^2

This requires changing the denominator from 2 * n_neighbors to 4 * n_neighbors:

kernel_matrix.data /= 4 * n_neighbors - kernel_matrix.data
kernel_matrix.data = kernel_matrix.data**2