Fix race condition where jobs are sent before workers register - backend now returns error for unavailable workers, cluster master re-queues jobs with delay

Add real-time GPU usage monitoring to cluster nodes - clients send GPU stats every 5 seconds, master stores and serves via API, web UI shows usage bars and utilization

Fix GPU memory and utilization stats to show actual values using pynvml instead of PyTorch-only stats

Fix backend to always return response for job forwarding - success/error messages ensure proper job assignment feedback

Fix progress message job_id to use integer ID instead of message ID for proper backend storage and web polling

Add progress-based timeout extension to prevent premature job timeouts during long-running operations