feat: add Prometheus metric for tracking unprocessed tasks

[vuonghp92]

Add a new Prometheus Gauge metric flower_unprocessed_tasks_in_window to track
tasks that have been processing (received but not completed) for longer than a
configurable time window.

Implementation details:

• Sliding window algorithm with time-based buckets for O(1) add/remove operations
• Configurable window size via --unprocessed_tasks_window_minutes option (any value > 0)
• Set to 0 (default) to disable tracking
• Buckets are timestamp-based integers for efficient lookups
• Counts bucket 2 steps back to ensure tasks have been unprocessed for > window_minutes
• Automatic cleanup of counted buckets to minimize memory usage

Data structures:

• time_buckets: Hash map of buckets containing task IDs
• task_to_bucket: Hash map for O(1) task-to-bucket lookup
• Both structures use dict for constant-time operations

Performance:

• add_task(): O(1)
• remove_task(): O(1)
• update_metrics(): O(n) where n = tasks in old bucket, runs periodically (once per window)

Use cases:

• Detect unprocessed or stalled tasks: Identify tasks that have been started but remain incomplete for longer than the configured time window, indicating potential processing delays or scheduling bottlenecks.

• Detect message loss in the broker or queue: Highlight scenarios where tasks disappear from the broker or queue system (e.g., RabbitMQ) but never reach the worker, suggesting possible message delivery or routing issues.

• Detect worker termination during processing: Monitor for tasks that remain “in-progress” because their worker was unexpectedly killed or crashed mid-execution, helping isolate stability or resource exhaustion problems.

Example:
flower --unprocessed_tasks_window_minutes=60