perf: export YOLO26 nano to ONNX or OpenVINO for the CPU pack

[Abhash-Chakraborty]

Summary

Keep YOLO26 nano as the quality baseline but evaluate exported CPU runtimes that avoid loading the full PyTorch/Ultralytics stack in the CPU pack.

Evidence

Find already configures yolo26n.pt, the current Ultralytics nano model. Official documentation reports YOLO26 supports ONNX, OpenVINO, LiteRT, CoreML, and other exports and improves CPU ONNX speed over YOLO11n: https://docs.ultralytics.com/models/yolo26

Where to work

• backend/src/find_api/ml/object_detector.py
• hardware provider resolution
• model manager/cache design in feat: design installer model downloads and cache management #45
• object-detection fixtures

Acceptance criteria

• Export and validate YOLO26n ONNX and OpenVINO CPU candidates.
• Measure artifact size, cold start, RAM, latency, and mAP/regression fixtures.
• Compare FP32, FP16 where supported, and INT8 with calibration documentation.
• Preserve the existing detected-object response contract.
• Add provider fallback and corrupt/missing-artifact tests.
• Recommend a CPU runtime only when quality loss stays within an agreed bound.

Out of scope

Do not replace YOLO26n with an older model merely because an export tutorial exists.

[Yuthika10]

Hi @Abhash-Chakraborty can you assign this issue to me?

[Abhash-Chakraborty]

@Yuthika10 how's this going? You have both this and #342 assigned since 11 July.

Two advanced/critical ML issues at once is a lot — if you'd rather focus on one, tell me which and I'll free the other up. No problem either way, I'd just rather it be deliberate than both stalling.