ONNX multi-exit export + end-to-end integration tests

[sohams25]

Removes the last shipped limitation (ONNX) and adds comprehensive isolation+integration testing.

ONNX export

• export_to_onnx / earlyon export write a static multi-output graph (one output per exit + final). ONNX has no portable per-sample control flow, so routing stays at runtime — the graph computes every exit and the caller picks the first confident one (honest trade: portability over the early-exit compute saving).
• Legacy tracer (dynamo=False) traces the static all-exits forward cleanly. Works for conv (4D) and transformer (3D-token) heads, with a dynamic batch axis. Saves/restores the model's train/eval mode (torch.onnx.export otherwise silently leaves it in train).
• onnxruntime added to the dev extra to verify exported graphs.

Testing — in isolation and together

• Isolation (tests/test_onnx.py): export cnn + token models, outputs match torch all-exits under onnxruntime, dynamic vs static batch, mode preservation, output-count tracks exits.
• Together (tests/test_integration.py): full pipelines — build → train → calibrate (confidence + entropy) → analyze → benchmark → save/load → ONNX — across a factory backbone (cifar_resnet) and a custom-wrapped transformer, plus a CUDA end-to-end run.

Test plan

• ruff · black · isort clean
• mypy --strict — 0 errors
• pytest -m "not gpu" — 159 passed, 97% coverage
• pytest -m gpu — 5 passed (RTX 4050)
• CI green on this PR

[sohams25]

Review — ✅ ready to merge

CI green across 3.10 / 3.11 / 3.12. This run installs onnx via the dev extra, so the ONNX tests genuinely run (not skip), validating export on CI's torch.

/review + a 4-dimension adversarial review found 11 findings (2 HIGH) — both HIGH were real ship-blockers and are fixed:

• dynamo=False kwarg only exists on torch>=2.4 (pyproject allows >=2.0): now applied only when the param is present (older torch defaults to the legacy exporter anyway), so export works across the whole supported torch range.
• onnx was an undeclared hard runtime dep (legacy exporter loads it; tests guarded on onnxruntime, so a clean runner would have failed, not skipped): now an [onnx] extra with a clear ImportError, added to dev, and tests importorskip("onnx").

Also addressed: deployment-contract test (apply the routing rule to ONNX outputs → matches torch inference, confidence + entropy), deprecation-warning suppression, dynamic-batch token coverage, narrowed the static-reject exception, eval-before-adapter, and the CLI now echoes the traced input shape.

Tested in isolation and together: 164 CPU + 5 GPU, 96% coverage. Merging.