- Fix cluster master to properly detect GPU backends from client capabilities
- Extract available_backends from capabilities list instead of gpu_info
- Ensure clients with GPU workers are correctly identified as GPU-enabled