feat: add kd_tree option for O(N log N) neighbor search in kNN estimators

[iPopovS]

What this does

Adds an optional kd_tree=False parameter to discover_network, conditional_mutual_information, knn_mutual_information, and geometric_knn_mutual_information. When set to True, neighbor searches use scipy.spatial.KDTree instead of a full pairwise distance matrix.

Why

The brute-force cdist approach is O(N²) in memory and time. KD-Tree queries are O(N log N), making a significant difference for large datasets.

Backward compatibility

kd_tree=False is the default, so all existing code is unaffected.

Testing

• Added tests/test_kdtree.py with correctness tests (kd_tree=True vs False produce numerically identical results) and a benchmark showing speedup.
• All existing tests continue to pass.

Benchmark results

N=300: brute=0.003s kd_tree=0.003s speedup=0.9x

N=800: brute=0.021s kd_tree=0.006s speedup=3.5x

Pre-existing test failures

TestGeometricKnnConditionalMutualInformation::test_geometric_knn_cmi_no_conditioning fails on the unmodified original code. The test compares geometric_knn_conditional_mutual_information(X, Y, Z=None, k=3) against geometric_knn_mutual_information(X, Y, k=1) — mismatched k values, so the assertion fails regardless of this PR's changes. Not introduced here.