fix: PG init backfill rollback scope + audit-loop state upsert registration

- db: translate INSERT OR IGNORE ... SELECT to ON CONFLICT DO NOTHING
  (VALUES-only regex left SELECT form untranslated -> PG syntax error);
  register task_audit_loop_state (task_id) in _OR_REPLACE_PK so PG
  INSERT OR REPLACE becomes a real upsert instead of DO NOTHING.
- server: _backfill_default_quotas uses SAVEPOINT containment and no
  longer commits the caller's transaction, so a backfill failure cannot
  roll back init_db's migrate ALTERs.
- worker: _coord_child_rows rolls back before column-fallback retries so
  a missing-column failure on PG does not cascade into
  InFailedSqlTransaction.
- tests: test_pg_init_fix.py (translation, SQLite backfill/upsert,
  real-PG fresh-init + merge + upsert + backfill containment, skipped
  without CLOUD_TEST_PG_DSN/driver).

cloud/worker.py

@@ -5772,6 +5772,14 @@ class Worker:
                         " WHERE id = ?", (cid,)).fetchone()
                     fetched.append((cid, row))
             except Exception:
+                # PG 下语句失败即 abort 整个事务：缺列的首次查询
+                # 失败后，不先 rollback 就重查只会连环报
+                # InFailedSqlTransaction。本连接只做 SELECT，rollback
+                # 无副作用；SQLite 下同样无害。
+                try:
+                    conn.rollback()
+                except Exception:
+                    pass
                 try:
                     fetched = []
                     for cid in child_ids:
@@ -5781,6 +5789,10 @@ class Worker:
                             " FROM tasks WHERE id = ?", (cid,)).fetchone()
                         fetched.append((cid, row))
                 except Exception:
+                    try:
+                        conn.rollback()
+                    except Exception:
+                        pass
                     fetched = []
                     for cid in child_ids:
                         row = conn.execute(