SaveSync: paginate recompute task scan by primary key

get_all_saves() materialized every Save row across all users into a
single .all() list. On instances with very large libraries that's a
real RAM ceiling and pins every row for the lifetime of the recompute
run.

Replace it with get_saves_after_id(after_id, limit) and have the
recompute task drive keyset pagination in PAGE_SIZE-row chunks. SQLAlchemy
streaming via .execution_options(yield_per=...) is incompatible with the
per-call session lifetime that @begin_session enforces (the session
exits before the consumer iterates), so keyset paging from the caller is
the cleanest fit.

Behavior is unchanged: same row coverage, same idempotency, same
counters. Memory usage drops from O(all saves) to O(PAGE_SIZE).

backend/handler/database/saves_handler.py

@@ -196,10 +196,18 @@ class DBSavesHandler(DBBaseHandler):
         )
 
     @begin_session
-    def get_all_saves(
+    def get_saves_after_id(
         self,
+        after_id: int,
+        limit: int,
         session: Session = None,  # type: ignore
     ) -> Sequence[Save]:
-        """Every Save row across all users, ordered by id. Used by the
-        recompute_save_content_hashes maintenance task."""
-        return session.scalars(select(Save).order_by(asc(Save.id))).all()
+        """Page Save rows by primary key. Returns up to ``limit`` rows with
+        ``id > after_id``, ordered by id. Used by the
+        recompute_save_content_hashes maintenance task to walk every row in
+        bounded-memory batches: streaming via ``yield_per`` is incompatible
+        with the per-call session lifetime that ``@begin_session`` enforces,
+        so the caller drives pagination with this method instead."""
+        return session.scalars(
+            select(Save).where(Save.id > after_id).order_by(asc(Save.id)).limit(limit)
+        ).all()