feat: add optional vision ONNX producer probe

[hyeokjun32]

Summary

• Added an optional ONNX Runtime probe for the device-local Vision producer.
• Exposed --vision-onnx-model for run-multi-workload-sustained.
• Recorded provider, input/output shapes, and probe latency in sustained telemetry without changing the default fixture path.

Validation

• PYTHONDONTWRITEBYTECODE=1 .venv/bin/python -m pytest tests/test_multi_workload_sustained.py -q -p no:cacheprovider
• PYTHONDONTWRITEBYTECODE=1 .venv/bin/python -m pytest tests/test_runtime_telemetry.py tests/test_config.py tests/test_workers.py -q -p no:cacheprovider
• PYTHONDONTWRITEBYTECODE=1 .venv/bin/python -m pytest -q -p no:cacheprovider
• PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=src .venv/bin/python -m inferedge_orchestrator run-multi-workload-sustained --config configs/agent_multi_workload_sustained_device_local.json --output /tmp/inferedge_orchestrator_device_local_onnx_probe_check.json --frames 4 --vision-input examples/inputs/vision_frame.ppm --vision-onnx-model /tmp/inferedge_orchestrator_identity_probe.onnx

Notes

• This is a lightweight device-local producer probe, not a full live YOLO service.
• Default sustained smoke remains dependency-free.