[python] Verify candidate-only scalar filters before vector top-k

[TheR1sing3un]

Purpose

Prevent candidate-only scalar index answers from admitting non-matching rows into vector top-k. For example, with a BTree name index and an IVF vector index, name LIKE '%zeta%' can currently return the nearer alpha row instead of beta zeta. Filtering after top-k cannot recover the displaced match.

Carry scalar-result exactness through offsets, unions, intersections, coverage padding and conjunctions that drop an unsupported child. BTree prefix/substring answers are candidates; exact BTree equality/range results and exact bitmap string matches retain their existing path without a data recheck. When multiple indexes answer the same predicate, an exact reader makes their intersection exact; conjunctions of different predicates still require every child to be exact. This preserves exact BTree-plus-bitmap matches without a data recheck. Tracking the result also handles partially evaluable conditions on the same field, where contributing field IDs alone are insufficient.

Use the global-index.filter.refine-from-data table option introduced by #9953, with the same default of false. Inexact index candidates are excluded with a warning by default. When enabled, read the candidate filter columns at the pinned snapshot to build an exact include bitmap before ANN top-k. This read is streamed, applies deletion handling, and does not project candidate vectors. Completely unsupported scalar evaluations can also use this opt-in verification. Local/Ray and single/batch vector execution share this path.

Related to #9883. This change covers Python data-evolution vector search; full-text filtering remains unchanged.

Tests

• New native-index and resource-lifetime tests: 53 passed, including all local/Ray, single/batch and full/fast combinations of the refinement option.
• The same 53 new tests also passed on Ray 2.44 / NumPy 1.24 / Arrow 18 (primary environment: Ray 2.54 / NumPy 2.4 / Arrow 19).
• Existing evaluator, union/sorted-index result and vector-filter regressions: 102 passed (155 total with the new tests).
• Tests cover BTree prefix/substring candidates, exact BTree and bitmap fast paths, mixed BTree-plus-bitmap filtering with refinement disabled/enabled, reader-order independence and conservative predicate conjunctions, unsupported predicates on the same field, result-wrapper propagation, historical snapshots/tags and deletions, a commit before candidate verification, projected columns and stream cleanup.
• Flake8, Python 3.6 syntax parsing and git diff --check passed for changed files. Native index tests used paimon-vindex 0.4.0.

[JingsongLi]

+1