Fix --purge-offline-worker-metrics

[guidoiaquinti]

I believe this PR fixes the --purge-offline-worker-metrics issue described at #362 so a late worker-offline event does not recreate a purged worker metric.

The bug could happen in this sequence:

1. A worker is purged, removing its metrics from the exporter registry.
2. A later worker-offline event arrives for the same hostname.
3. track_worker_status() recreated celery_worker_up{hostname="old-worker"} 0.0.
4. Since the worker had already been removed from worker_last_seen, the recreated zero-valued series was no longer eligible for future purge.

Now only online events directly create/update celery_worker_up. Offline events go through forget_worker(), which already no-ops for unknown or already-purged workers.

Note: AI assistance was used during investigation and validation.