Make scikit-learn an optional dependency

Goal

Turn scikit-learn (and skl2onnx) into an optional dependency, mirroring how PyTorch support works: users who only need parse_pytorch_model can install orbital without pulling scikit-learn/skl2onnx.

Done when:

• A PyTorch-only install does not include scikit-learn/skl2onnx.
• import orbital works without scikit-learn installed; calling parse_pipeline (or other sklearn-only entry points) without it raises a clear ImportError pointing to the extra to install (e.g. pip install orbital[sklearn]).
• Decide and document what a bare pip install orbital provides (keep sklearn for backward compatibility vs. fully optional extras for both frameworks).

Why

PyTorch users currently pay for dependencies they never use: scikit-learn, skl2onnx, and packaging are unconditional core deps. Symmetric treatment of both frameworks keeps installs lean and the dependency story coherent.

References

• pyproject.toml (dependencies: scikit-learn, skl2onnx, packaging — the latter noted as required by skl2onnx converters)
• src/orbital/ast.py (module-level skl2onnx/sklearn.pipeline imports, parse_pipeline)
• src/orbital/types.py (wraps skl2onnx.common.data_types — types are shared by all parse paths, so this import is the hard knot)
• Branch nnet (not yet on main): parse_pytorch_model lazy-import pattern and the pytorch extra — the model to replicate