Use per-dimension mean (axis=0) in mahalanobis metric

[uttam12331]

Summary

The Mahalanobis distance centers X on the mean vector (per-column centroid) of the reference distribution, but _compute uses np.mean(reference_distribution) with no axis, which collapses the 2-D reference array to a single scalar grand mean and broadcasts it across every feature:

X_minus_mu = X - np.mean(reference_distribution)      # scalar grand mean
cov = np.cov(reference_distribution.T)                # per-column covariance

The very next line, np.cov(reference_distribution.T), treats each column as a variable — confirming the per-dimension intent. As a result the metric is wrong for any distribution whose feature means differ.

Reproduction

import evaluate
m = evaluate.load("mahalanobis")
m.compute(reference_distribution=[[0, 10], [4, 10], [0, 14], [4, 14]], X=[[2, 12]])
# X = [2, 12] is exactly the centroid, so the distance should be 0.0
# actual:   {'mahalanobis': array([9.375])}
# expected: {'mahalanobis': array([0.])}

Fix

-        X_minus_mu = X - np.mean(reference_distribution)
+        X_minus_mu = X - np.mean(reference_distribution, axis=0)

Tests

Added a doctest showing a point on the distribution's centroid has distance 0. The existing doctest (reference_distribution=[[0, 1], [1, 0]]) is symmetric — its column means equal the grand mean — so it yields array([0.5]) under both the old and new code and remains unchanged.