Feed drops are classified without needing a configured LLM

Feed ingestion no longer depends on having a configured large language model. The GitHub issue, GitHub release, and RSS feed syncers now run as Oban workers scheduled through Cron, with Oban started unconditionally in the supervision tree before the endpoint. Drops keep being created and classified using the deterministic fallback path, and the manual feed sync button in the dashboard is routed through the same worker, so background and on-demand syncs are reliable even on instances where model-backed agents are dormant.