fix(cytovi): stabilize log-median probability aggregation

[sofoklis326]

CytoVI's log_median() currently computes:

np.log(np.median(np.exp(x), axis=axis))

The input already contains log probabilities. Exponentiating very
negative values underflows to zero, while very positive values overflow
to infinity. This can produce non-finite differential-abundance scores.

For example, groups with constant log scores of -900 and -880 should
produce a DA difference of -20. The current calculation instead produces
-inf - (-inf), yielding NaN.

This PR selects the central values directly in log space and combines
them using np.logaddexp. It preserves the probability-space median
definition for both odd and even sample counts.

Compatibility checks cover aggregation axes, missing values, infinities,
empty reductions, and preservation of the input array.

Validation:

• Added 37 focused test cases.
• On unmodified upstream, 16 cases fail and 21 pass.
• With the fix, all 37 cases pass.
• The existing CytoVI training smoke test also passes locally on CPU:
  38 passing tests in total.
• Focused Ruff and formatting checks pass.
• Added a changelog entry.

This fix is independent of the other CytoVI fixes proposed separately.

The full upstream CI matrix has not been run locally.

[ori-kron-wis]

was solved here: #4048