[fix](inverted index) Exclude MoW light deletes from BM25 statistics - #69403

What problem does this PR solve?

Issue Number: None

Related PR: #69352, #69438

MoW light deletes use versioned DELETE predicates. The storage reader filters their rows, but BM25 collection statistics only exclude delete-bitmap and delete-sign rows. The same visible documents therefore score differently from an independently loaded MoW table before compaction.

Extend the existing tablet-wide statistics collection with the captured DELETE predicates. Reuse DeleteHandler for current and historical column identities, compound/NULL predicate semantics and the query/rowset version boundaries. Read only the referenced columns in bounded batches and union filtered row IDs with the existing hidden set. The existing document-count, term-frequency and norms-sum corrections then apply unchanged to CLucene and SNII, including prefix expansion.

This keeps the existing shared-statistics mechanism, per-scanner derived caches and MoW norms-based avgdl. It adds no exact-length cache or raw-text tokenization. A newer insert is not hidden by an older DELETE. Scans with no DELETE predicates require no additional column reads.

Scope: MoW with its normal delete bitmap. DUP/MOR tables and the independent skip_delete_sign/show_hidden_columns issue are outside this change. DUP currently uses exact token totals, so extending its visibility statistics requires a separate decision about preserving that scoring contract; it must not silently switch to norms only while predicates exist.

Release note

Correct BM25 statistics on merge-on-write tables after light deletes, without waiting for compaction.

Check List (For Author)

• 93 selected ASAN tests passed in CollectionStatisticsTest, CollectionStatisticsDetailedTest, CollectionStatisticsHiddenRowsTest, TabletCollectionStatisticsTest and ScoringDeletePredicatesTest. The three new real-segment cases cover batch boundaries, overlap with bitmap rows, historical column UIDs and query/rowset version bounds.
• Built the master FE and the patched ASAN BE through build.sh, using baseline 8e10b23. No FE source or persisted format changes; optional BE Java extensions disabled locally.
• test_bm25_score_light_delete_statistics passes on that cluster: V2/V3 with support_phrase=true/false and SNII with support_phrase=true; compound and NULL predicates, inserts after a delete, overlapping overwrite/delete histories, all-hidden/refill and serial/forced-parallel scans.
• Existing test_bm25_score_parallel_scan_statistics and test_bm25_score_replica_statistics pass on the same cluster. This is a single-BE test cluster; the replica suite's multi-replica-only steps are not exercised locally.
• The same new regression fails on the uncorrected query path at its first score comparison. This was also reproduced on a freshly built pure-master ASAN BE at the same frozen baseline and with the same master FE.
• clang-format 16, build hygiene, and changed-line clang-tidy on all modified C++ source/test code passed.
• Behavior changed: Yes. Light-deleted rows no longer affect MoW BM25 scores.
• Does this need documentation: No new configuration or syntax.

Cost and remaining scope

The additional work reads the predicate columns of segments to which a captured light-delete predicate applies. It runs once per tablet through the existing holder. Added row-ID arrays use DorisVector and are disjoint from bitmap/sign rows before a linear merge. There is no new permanent cache, but query work and temporary memory grow with affected stored rows; no production performance claim is made.