Purge offline workers and clean up Prometheus metric labels

[ShubhAtWork]

Summary

When a Celery worker goes offline permanently (e.g. a scaled-down container, a crashed node, a spot instance termination), its state persists indefinitely in three places:

1. EventsState.counter — per-worker event counter dict (~1 KB per worker)
2. Inspector.workers — cached inspect results
3. Prometheus label series — every labels(worker_name, ...) call creates a time series that persists for the lifetime of the process

In long-running deployments with worker churn (autoscaling, rolling deploys, spot instances), these accumulate without bound, causing monotonically growing memory usage, degraded Prometheus scrape performance, and a slow /workers page.

Problem with the existing --purge_offline_workers option

The option already exists in flower/options.py, and WorkersView.get() uses it to filter workers from the UI. But it does NOT actually remove them from memory — the counter, inspector, and Prometheus data continue to grow.

What this PR does

• Add a periodic timer (runs every purge_offline_workers seconds, minimum 10s) that scans for workers offline beyond the configured threshold
• Actually purge dead workers from state.counter, inspector.workers, and all Prometheus metric label series
• Handle orphaned entries — workers that appear in counter/inspector but not in state.workers (e.g. from incomplete event sequences)
• Add PrometheusMetrics.remove_worker_metrics() — iterates all 6 metric families and removes label series matching the worker name
• Add Inspector.purge_worker() — removes cached inspect data for a worker

Test plan

• pytest tests/unit/test_app.py tests/unit/test_events.py tests/unit/test_inspector.py — 14 tests pass
• Deploy with --purge_offline_workers=600, verify memory stabilizes after workers scale down
• Verify Prometheus /metrics endpoint no longer shows labels for purged workers

🤖 Generated with Claude Code