fix(search): make indexing converge and survive shutdown

[zoom-ua]

adapted from #29467

@mmaxim

[zoom-ua]

1. Evicted dirty entries lost — Entries evicted from LRU cache before flush were permanently dropped; now tracked in pending set and written on next flush
2. Failed writes dropped — Flush cleared dirty tracking before writing to disk; write failures left messages marked indexed with no tokens (unsearchable forever); now requeues failed writes for retry
3. Deletes racing with flush — Direct-to-disk deletes raced with concurrent flush snapshots, resurrecting deleted entries; now queues tombstones that flush applies in order
4. Metadata published ahead of tokens — Metadata (SeenIDs) written directly while tokens were still pending; convs read as complete with missing tokens; now routes all writes through overlay
5. Reindexing refetches everything — Paged over full min..max range instead of missing IDs only; spent budget refetching already-indexed messages; now pages over missing IDs
6. Deleted/expired messages block convergence — IDs the server won't return (deleted messages, retention policy expiry, gaps) held numMissing > 0 forever, causing infinite refetch loops; now marks unreturned IDs seen after successful fetch
7. Stop/Start lifecycle races — Concurrent Stop/Start could replace store while old run still uses it; now serializes runs (Start waits for previous Stop to finish)
8. Storage failures unreported — Add/Remove ops always completed callback successfully even when indexing failed, causing reindexConv to mark unindexed messages as done; now propagates errors via callback
9. Clear() global side effects — Clearing one conv discarded all other convs' pending writes (ClearMemory is global, disk clear is per-conv); now flushes before clearing