feat: export classifier and regressor pipelines to ONNX

[serhiizghama]

Closes #307.

scripts/export_to_onnx.py could only export the encoder (StaticModel) — there was no way to get a trained StaticModelForClassification/Regression pipeline into ONNX, which is what the issue asks for. This teaches the same script to handle both: it auto-detects whether the given path is a plain encoder or a pipeline with a head, and exports accordingly, so the existing --model_path/--save_path interface is unchanged.

The classifier path wraps the encoder plus the MLP head into one torch.nn.Module. The head's Layers store weights as [out, in] and compute x @ weight.T + bias, which is exactly nn.Linear(in, out), so rebuilding them is a direct copy. The output activation matches the head: softmax/sigmoid produce probabilities, an identity head (regressor/projector) produces raw predictions, and the ONNX output is named accordingly.

Added tests that export a trained pipeline and run it through onnxruntime, checking the result against predict_proba for classifiers and predict for the identity head — single-label, multilabel, and projector cases all match within 1e-4. Existing inference tests still pass, and ruff/mypy are clean.

[stephantul]

Hey this is cool, thanks for making it. We were about to add this ourselves, but nice you got to it first. I'll check it out ASAP

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[stephantul]

Thanks for adding this! I will make some changes to the exporting code in parallel, so this is fine to merge.