Add opt-in layered BM25 search over clustered FTS terms

[kryonix]

This PR adds an opt-in layered_search mode to the FTS extension. The new path keeps the existing FTS occurrence storage, but adds a small dictionary-level sidecar for query-time expansion and BM25 scoring over clustered terms.

• Adds layered_search=true to PRAGMA create_fts_index.
• Makes layered_search imply cluster_terms, so terms is physically ordered by termid, fieldid, docid.
• Builds extension-owned dictionary sidecar tables:
  • term_stats
  • term_stats_by_len
  • term_grams
• Adds layered search macros:
  • search_layered_bm25
  • match_layered_bm25
• Supports exact, prefix, substring, and fuzzy dictionary expansion.
• Keeps numeric terms exact-searchable, but excludes them from trigram expansion.
• Keeps this path postings-free; scoring still uses the existing FTS terms table.

When incremental=true is used with layered_search=true, the sidecar is maintained by triggers for inserts and deletes. The triggers update only affected terms derived from the statement's new/old rows, instead of refreshing the whole sidecar.

Deletes also prune zero-df terms from the layered sidecar so removed vocabulary no longer contributes to query expansion.

[Dtenwolde]

Codex identified an issue with the configured stopword filter being applied before expansion. Normally these are filtered out before stemming and before they enter dict/term_stats. But search_layered_bm25 only removes NULL and empty strings:

WHERE query_term IS NOT NULL
        AND query_term <> ''

Therefore, a stopword-only query such as the can slip through this and fuzzy-expand into terms like theme (eventhough the was supposed to be excluded from the index).

We should probably filter raw query tokens against the stopwords table before stemming/expansion (except when stopwords='none').

[kryonix]

Thanks @Dtenwolde I fixed it. But main moved again, so I need to update the submodule first for CI.

[lnkuiper]

I've merged the duckdb pin PR, so this should update the submodule