fix: scheduler performance and correctness improvements

[deepjoy]

• Replace tokio::Mutex with std::sync::Mutex in ActiveTaskMap — most operations do trivial HashMap work with no .await, so a sync mutex avoids unnecessary runtime interaction. Methods that need async I/O (preempt_below, pause_all) now use two-phase locking: collect under the sync lock, then release before awaiting.
• Eliminate O(N²) progress fan-out — remove the per-task broadcast listener that caused O(N×M) wakeups across N tasks and M events. ProgressReporter now updates the ActiveTaskMap directly via a sync call.
• Break Arc cycle in SpawnContext — replace the strong Scheduler reference with WeakScheduler to prevent keeping SchedulerInner alive during graceful shutdown.
• Track JoinHandles in ActiveTask — shutdown can now join spawned futures instead of polling active.count() in a 50ms busy loop. Graceful shutdown joins each handle with a per-task timeout budget.
• Atomize single-task submit() — wrap in BEGIN IMMEDIATE transaction to match submit_batch() and prevent non-atomic multi-statement races.
• Wake scheduler after task completion and failure — add missing work_notify.notify_one() calls after Completed and Failed event emissions so the scheduler loop dispatches pending work immediately instead of waiting for the next poll interval.