Add decoded-output parity gates for Vision v8 ONNX exports

[Complexity-ML]

Context

PR #10 calibrated branch-specific raw-logit tolerances for Vision v8 because CPU ONNX Runtime drift is concentrated in fine-grid regression logits while decoded box and class-score drift remains much smaller.

Raw-logit tolerance is still only a proxy for deployment behavior. The validation report explicitly recommends checking decoded outputs.

Scope

• Extend scripts/check_onnx_parity.py to compare decoded outputs in addition to raw logits.
• Decode LTRB/DFL regression using the exported model metadata and detector grid configuration.
• Compare normalized boxes and sigmoid quality-class scores for both o2m and nms-free.
• Report raw-logit, decoded-box, and decoded-score max/mean differences separately.
• Keep deterministic inputs and make each gate fail independently.
• Add tests for both branches, including a case where raw drift is tolerated but decoded drift exceeds its limit.

Acceptance criteria

• Both v8 branches have documented decoded parity thresholds with numerical justification.
• The CLI output clearly identifies which parity gate failed.
• Existing legacy exports retain their current strict raw-logit behavior when decoded metadata is unavailable.
• tests/test_detector_export.py covers the new behavior.
• The detector-export workflow passes on CPU ONNX Runtime.

Related: #10

[IIIllllIlIlllII]

Working on this.

Recalibrated the decoded thresholds on the SFT checkpoint over 50 seeds. Note that the calibration also turned up a pre-existing issue: the v8 raw thresholds from #10 were fitted on 5 seeds and are exceeded at 20+. PR incoming.