fix(fts): skip rows COPY rejects with ignore_errors

[zahariash]

This work was produced with the help of language models.

Fixes #118.

COPY ... (ignore_errors=true) into a table with an FTS index mishandled a row rejected for a duplicate primary key. COPY indexes its rows in FTSIndex::finalize, but the rejected row was deleted before that, so the index looked for a document it had never seen and the COPY failed. A rejected row without text instead became a phantom document with the previous row's text, skewing BM25 scores.

Fix. delete_ skips a row at or past the indexed prefix (checkpointed rows plus this transaction's inserts) when the docs table has no document for it, so an indexed row is still deleted. finalize checks lookup() and skips deleted rows instead of indexing the previous row's vectors. This also makes rows a crash left outside the index deletable again, until a later insert moves the prefix past them; they stay unsearchable, tracked by LadybugDB/ladybug#1241.

Tests. copy_ignore_errors.test copies a duplicate with text and checks the other row is indexed and deletable, then one without text and checks the score equals a rebuilt index's, and one where the rejected row is the last in the file, followed by an insert and deletes. The new cases fail on main and pass here. FTS extension tests: no new failures.

[adsharma]

@zahariash the exact score match may be a bit brittle and require frequent updates due to code/data changes.

[zahariash]

Fixed in 38c08a2: the pinned score is gone. The test now builds one index before the COPY (maintained through it, so it would pick up the phantom document) and one after (which can't), and checks both score the document the same. Still fails with 2|False without the fix.

[zahariash]

38c08a2 went onto the branch after the merge, so it never reached main. The same change is #126.