Add decoded-output parity gates for Vision v8 ONNX exports

[IIIllllIlIlllII]

Closes #12.

scripts/check_onnx_parity.py now evaluates three independent gates on the same
deterministic inputs — raw, decoded-box (normalized xyxy after LTRB/DFL
decode) and decoded-score (sigmoid quality-class) — each with its own max/mean
report, its own pass/fail, and its name printed in the summary when it fails.
Decoding goes through the complexity.deploy.onnx_detector package merged in
#14, so the gates exercise the deployment path rather than a parallel
implementation.

Why raw tolerance cannot stand in for decoded behaviour

Two effects pull in opposite directions. Softmax is translation-invariant, so a
drift shared by all 17 DFL bins of a distance moves the decoded box by exactly
zero. Conversely the decoded distance is multiplied by the stride, so a small
drift on the 20x20 grid moves a box by several pixels. Raw and decoded drift are
therefore not monotonically related, and each gate has to stand alone.

Please read this part before the diff: the raw thresholds are loosened

The first commit raises the v8 raw thresholds (o2m 2e-3 → 6e-3, nms-free
3.5e-3 → 1e-2). That is not to make anything in this PR pass — nothing here
touches raw drift, which still measures 1.74e-03 / 2.99e-03 at five seeds as
in #10.

The reason is that calibrating on five seeds underestimated the maximum. Each
test compares ~5M values, so the observed max is an extreme-value statistic that
keeps growing with the seed count:

Branch	max@5	max@50	growth	old threshold
o2m	1.741886e-03	2.788305e-03	+60.1%	2.0e-03
nms-free	2.990723e-03	4.793882e-03	+60.3%	3.5e-03

Both branches exceed their threshold well before 50 seeds, so
check_onnx_parity.py --num-tests 20 fails on main today on a healthy export.
Reproduce with --num-tests 5 then --num-tests 50 and compare the reported
worst_max.

Decoded outputs grow only +12% to +34% over the same range — three to five
times more stable — which is itself the quantitative argument for gating on
them. All thresholds now follow one rule: twice the maximum observed over 50
deterministic seeds.

Branch	Gate	max@50	threshold	headroom
o2m	raw	2.788305e-03	6.0e-03	2.2x
o2m	decoded-box	5.497933e-05	1.3e-04	2.4x
o2m	decoded-score	3.811717e-05	8.0e-05	2.1x
nms-free	raw	4.793882e-03	1.0e-02	2.1x
nms-free	decoded-box	6.532669e-05	1.3e-04	2.0x
nms-free	decoded-score	1.719594e-05	4.0e-05	2.3x

decoded-box is shared across branches (box drift differs by 19%, same
regression head); decoded-score is branch-specific (o2m drift is 2.2x
nms-free).

One correction to the report

The "decoded boxes max diff" figures in the validation report were in normalized
coordinates, which the report did not state. The unit is now explicit, and the
values are recomputed with the shared decoder, which is why they shift slightly.

Legacy behaviour

Sidecars without v8 decode metadata skip the decoded gates rather than failing
them, and keep the strict 1e-4 raw threshold. --skip-decoded forces that on
any export.

Tests

tests/test_detector_export.py covers both branches, the legacy fallbacks, and
the amplification case #12 asks for: tilting the DFL logits by epsilon * bin
defeats softmax cancellation, so raw drift stays at 4e-03 under a 5e-03
tolerance while normalized box drift reaches 5e-04 against a 1.3e-04 one —
and decoded-score stays green, demonstrating gate independence.

Measured on AETHORIA-AI/TR-HASH-Vision-v8-2M-COCO-SFT with PyTorch
2.13.0+cu130, ONNX 1.21.0, ONNX Runtime 1.24.4, CPUExecutionProvider.
Raw maxima reproduce #10's to three significant digits despite the different
runtime versions.