Add an iLISI and silhouette equivalence test against scib-metrics

mantispy.metrics reimplements scib's definitions rather than depending on scib, and tests/test_api_guards.py asserts that src/ never imports scib, scib-metrics, pycytominer, pyod, cytominer-eval or scmorph, so an equivalence check cannot become circular.

The gap: nothing actually checks the reimplementation against a reference. tests/test_metrics.py used to claim "equivalence with scib where the definition is shared" in its module docstring while no test performed one; #51 corrects the docstring rather than adding the test.

Worth adding as a test-time-only dependency, in the spirit of the existing pycytominer and scmorph equivalence tests: compare iLISI, cLISI and the label/batch silhouettes against scib-metrics on a fixed synthetic object, asserting numerical agreement rather than a conceptual match. That would earn back the sentence that was removed.

Note the LISI implementation already claims agreement with harmonypy.lisi.compute_lisi to machine precision; a scib-metrics comparison would cover the silhouettes and PC regression too.

Start with tests/test_metrics.py and the existing pycytominer and scmorph equivalence tests, then review tests/test_api_guards.py to preserve the no-source-import rule. Add scib-metrics only for testing and compare iLISI, cLISI, label silhouettes, batch silhouettes, and PC regression on a fixed synthetic object. Done means the numerical comparisons pass and the API guard remains satisfied.