fix(fts): tolerate interrupted CREATE/DROP residue (fixes #77)

[adsharma]

Fixes #77.

CREATE_FTS_INDEX / DROP_FTS_INDEX execute as sequential statements outside a single transaction (COPY can't run in a manual transaction), so anything failing mid-way left residue with no in-place recovery: next CREATE failed already exists, a torn DROP could reach a state no command accepts, and a torn index made the whole database fail to open at LOAD.

Changes (all localized to create_fts_index.cpp / drop_fts_index.cpp / fts_extension.cpp):

• Self-healing CREATE: when the index entry is absent, schema-confirmed residue (appears_in/appears_info/docs/terms, per-index stopwords copy) is dropped first via DROP TABLE IF EXISTS, so a failed CREATE is simply retryable. Two safeguards so user data is never dropped: (1) tables visible in the regular catalog still raise; (2) internal-set candidates are dropped only if their schema exactly matches the FTS internal schema (node shapes incl. PK, rel shape incl. endpoints when resolvable) — anything else raises. The shared default_english_stopwords table is never dropped.
• Flag-leak guard: useInternalCatalogEntry is save/restored around cleanup because ClientContext resets it only on the success path — without this, a failed CREATE misfiles later user DDL into the internal set (candidate engine follow-up in LadybugDB/ladybug).
• Guarded DROP: all table drops use IF EXISTS in referential order (rel first) and now also cover appears_info and the per-index stopwords copy (resolved from index aux info at rewrite time), making an interrupted DROP retryable. An entry-gone/tables-remain state is recoverable via re-CREATE.
• Tolerant LOAD: initFTSEntries pre-checks all four internal tables + stopwords; an incomplete index is skipped instead of throwing, so the database still opens. Repair with DROP_FTS_INDEX + CREATE_FTS_INDEX.
• Out of scope: tokenize-macro cleanup. Emitting DROP MACRO triggers a pre-existing engine bug (macro-drop WAL replay looks up the wrong catalog set and breaks un-checkpointed reopen — reproducible with plain CREATE/DROP MACRO SQL, no FTS involved). Macros keep the old skip-if-exists reuse behavior; revisit once the engine bug is fixed.

Verification: release build; full local FTS e2e suite green (except env-dependent show_loaded_extensions, an artifact of the partial local build); real kill -9 mid-CREATE (300k docs) → reopen fine → retry CREATE succeeds → queries correct. Remaining CI matrix left to CI.

[zahariash]

LadybugDB/ladybug#1165 is on main, so the macro cleanup deferred here is unblocked. #124 does it (drops the macro in DROP_FTS_INDEX, replaces it in CREATE_FTS_INDEX, tests in drop_recreate.test) and gets rebased onto this once it lands, which also removes the NOTE in drop_fts_index.cpp; its own stopwords handling is redundant with this PR's and goes away in that rebase.