fix: restore Poisson CMI sign (follow-up to #42)

[Parva-26]

Follow-up to #42, as discussed in that thread.

Problem

#42 changed the last line of poisson_conditional_mutual_information to cmi = H_YZ - H_XYZ. H_YZ and H_XYZ are differences of poisson_joint_entropy terms built from the adjusted matrix SS, not plain conditional entropies, so with that sign the raw CMI is negative for dependent data. The wrapper clamps it to 0 and Poisson discovery finds no edges.

Changes

• Restore cmi = H_XYZ - H_YZ (the order used by Compute_CMI_Poisson in jefish003/NetworkInference) and replace the inline comment that suggested the opposite. The np.ix_ indexing from fix: kde_entropy numerical stability, np.matrix removal, Poisson CMI sign #42 is kept.
• test_poisson_cmi_via_main_function: compare the wrapper with max(0.0, direct). X, Y and Z are independent in that test, so the raw value can be slightly negative and the wrapper clamps it. The test only passed before because of the old indexing offset and then because of the flipped sign.
• test_poisson_cmi_count_data_properties: also check that the conditional estimate is positive (it is -0.66 on current main).
• New test_poisson_cmi_sign_dependent_vs_independent: over 20 datasets, the raw CMI is positive on every dependent dataset (mean > 0.2) and its mean is close to 0 on independent data.
• New causationentropy/tests/test_poisson_discovery.py: a ~4 second standard oCSE run with information="poisson" on poisson_coupled_oscillators, so the Poisson path is covered by the default suite. The existing Poisson integration tests are deselected by default, which is why fix: kde_entropy numerical stability, np.matrix removal, Poisson CMI sign #42 passed CI.

Validation

• Default suite: 465 passed, 17 deselected. The three new or changed Poisson checks fail on current main.
• python -m pytest causationentropy/tests/test_data_integration.py -k poisson -o addopts="": test_standard_poisson and test_alternative_poisson both pass (they fail on current main, as reported in fix: kde_entropy numerical stability, np.matrix removal, Poisson CMI sign #42).
• The discovery script from fix: kde_entropy numerical stability, np.matrix removal, Poisson CMI sign #42 (n=5, T=300, seeds 0-2): 8/8 true edges with this change, 0/8 on current main.
• black --check, isort --check-only and the flake8 hard gate pass.

Not included: the SS = SXYZ aliasing and fill_diagonal questions that @jefish003 is looking into.