fix(database): batch database saves for bulk host writes

[ZacharyZcR]

On SQLite every repository write force-saves: the in-memory database is serialized and encrypted into fresh buffers and written to disk. 2.9.0 moved per-host options into plugin settings, so the boot plugin data migrations, host defaults materialization (run at boot and after every host write) and the bulk host routes now write one row per host per setting in awaited loops. Each loop rewrote the full database hundreds or thousands of times, and because the loops never yield, V8 could not free the copies in between. The #1380 log shows the container being OOM-killed right after Loaded 34 plugin(s), which is exactly where runPluginDataMigrations runs. In a standalone repro (60 MB database, the same serialize + AES-GCM + write as saveMemoryDatabaseToFile), 100 back-to-back saves peaked at 6.4 GB RSS.

Two changes:

• DatabaseSaveTrigger.batched(fn) wraps a function so the force saves made while it runs collapse into one save when it finishes (an AsyncLocalStorage scope, so concurrent requests outside it still save immediately). Nested scopes fold into the outer one, and work that outlives its scope (fire-and-forget) saves normally again. Applied to runPluginDataMigrations, materializeHosts, and the /bulk-import, /ssh-config-import and /bulk-update handlers. A 50-host import now saves once instead of once per row.
• runSave yields one event-loop turn after each save before resolving, so any remaining write loop frees each save's copies before the next one starts (369 MB peak instead of 6.4 GB in the repro above).

/database/import is left as is: it already ends with an explicit save whose failure it reports to the caller. For the wrapped routes the single save runs right after the handler returns rather than after each row.

Fixes Termix-SSH/Support#1384.
Fixes Termix-SSH/Support#1380.

Validation: database-save-trigger.test.ts covers collapsing, nested scopes, empty batches, failed batches and work that outlives its scope, plus the event-loop yield; full vitest run src/backend shows no new failures against the base (remaining failures are pre-existing local environment ones); eslint, prettier 3.9.6 and tsc -p tsconfig.node.json pass.