[Python] Deflake anomaly transform tests

[bvolpato]

Fixes #37779.

RunOneDetector reshuffles input before scoring. Existing transform tests used online Z-score detectors and matched each example against scores calculated in creation order. Runner scheduling can legally change that order independently for each ensemble branch, changing detector state and scores. This caused the keyed failure reported in #37779 and a new unkeyed failure in Python ML precommit.

Replace online Z-score use only in the ensemble tests with a small value detector and deterministic score fields. Original x1 and x2 data, expected predictions, and the single-detector Z-score test remain unchanged. Coverage remains for keyed and unkeyed inputs, feature selection, thresholds, ensemble grouping, and AnyVote aggregation.

Reproduction

Forcing one detector branch to consume reversed input reproduces score mismatches before this change. Deterministic score fields remove branch-order dependence from the four ensemble cases.

Testing

• ./gradlew :sdks:python:test-suites:tox:py313:testPython313 -PpythonVersion=3.13 -Pposargs='apache_beam.ml.anomaly.transforms_test' (12 passed)
• Four ensemble cases repeated 20 times with two pytest workers (80 passed)
• YAPF, Ruff, and Pyrefly checks

No CHANGES.md entry. Test-only change.