Backfill premature completion (confirmed on Airflow 3.2.2)

### Under which category would you file this issue? Airflow Core ### Apache Airflow version 3.2.2 ### What happened and how to reproduce it? Again more slop but hopefully useful enough. `<🤖>` --- When a backfill is created for a DAG with fast-completing tasks (sub-second per run), the scheduler marks the backfill as complete before all queued runs have been executed. The root cause is in `_mark_backfills_complete` (`scheduler_job_runner.py` ~line 1967), which runs every 30 seconds and marks a backfill complete when no dag runs are in `running` or `queued` state: ```python ~exists( select(DagRun.id).where( and_(DagRun.backfill_id == Backfill.id, DagRun.state.in_(unfinished_states)) ) ) ``` When tasks complete faster than the scheduler's next scheduling loop can queue new runs, there is a window where all current runs are `success` and the next batch has not yet been dispatched. The completion check fires in this window and incorrectly marks the backfill done, leaving remaining queued runs stranded. **To reproduce:** 1. Create a DAG with a no-op task and a long date range: ```python from airflow.sdk import dag, task from datetime import datetime @dag( dag_id="test_backfill_bug", schedule="@daily", start_date=datetime(2020, 1, 1), end_date=datetime(2022, 12, 31), catchup=False, ) def test_backfill_bug(): @task def noop(): pass noop() test_backfill_bug() ``` 2. Create a backfill: ```bash airflow backfill create \ --dag-id test_backfill_bug \ --from-date 2020-01-01 \ --to-date 2022-12-31 \ --max-active-runs 10 ``` 3. Observe that the backfill completes having only processed a fraction of the 1096 runs: ```python import sqlite3, os conn = sqlite3.connect(os.path.expanduser('~/airflow/airflow.db')) cur = conn.cursor() cur.execute('SELECT id, completed_at FROM backfill WHERE dag_id="test_backfill_bug"') b_id, completed_at = cur.fetchone() cur.execute('SELECT state, COUNT(*) FROM dag_run WHERE backfill_id=? GRO…