fix(kk): classify wide abbreviations per occurrence

[yisding]

Motivation

• The Kazakh refactor moved formerly-dotted abbreviation protection into the period-classifier global-realize model, but the classifier deduplicates candidates by (abbreviation, follower_char) and can drop later same-key occurrences that need different decisions, producing incorrect splits in cases like обл.x обл.: қала.

Description

• Enable per-occurrence realization for the Kazakh policy by setting realize_per_occurrence=True on KK_POLICY in sentencesplit/lang/kazakh.py so each candidate is classified from its own original context.
• Add a regression test test_kk_empty_follower_occurrences_are_classified_independently in tests/lang/test_kazakh.py asserting the обл.x обл.: қала example segments as ['обл.', 'x обл.: қала'].
• The change preserves the widened follower-class behavior for the frozen set of formerly-dotted stems via the existing regular_follower_overrides field and only alters realization semantics.

Testing

• Ran the focused Kazakh and classifier tests with uv run pytest tests/lang/test_kazakh.py tests/regression/test_abbr_dot_normalization.py tests/contract/test_period_classifier.py and they passed (all tests in that run succeeded).
• Ran lint/format checks with uv run ruff check . and uv run ruff format --check . and they succeeded.
• Ran the full test suite with uv run pytest --cov=sentencesplit tests/ --color yes and the test run completed successfully (the suite passed with the repository's existing skips/xfails).

---

Codex Task

[charliecreates]

Reviewed the latest changes and I don’t have actionable feedback to address. CI is still in progress, so merge readiness depends on those checks completing.