Split Python publishing: nlpo3 (base) + nlpo3-deepcut (optional ONNX package)

[Copilot]

Bundling the deepcut ONNX model and tract-onnx runtime into every nlpo3 wheel added ~7 MB to an install most users don't need. This splits the Python publication into a lightweight base and an opt-in deepcut package.

New packages

Package	Directory	Wheel size	Contents
nlpo3	nlpo3-python/	~1.1 MB	NewmmTokenizer, NewmmFstTokenizer, deepcut shim
nlpo3-deepcut	nlpo3-deepcut-python/	~8.5 MB	DeepcutTokenizer + ONNX model + tract-onnx

Base package (nlpo3)

• Compiled without deepcut Cargo feature → no tract-onnx, no ndarray, no embedded ONNX bytes
• DeepcutTokenizer remains importable from nlpo3 as a Python-level forwarding shim:
  • Present: nlpo3_deepcut.DeepcutTokenizer is constructed and returned
  • Absent: raises ImportError with install hint

from nlpo3 import DeepcutTokenizer
tok = DeepcutTokenizer()  # works if nlpo3-deepcut installed; else:
# ImportError: DeepcutTokenizer requires the nlpo3-deepcut package.
# Install it with:  pip install nlpo3-deepcut

New nlpo3-deepcut package

• nlpo3-deepcut-python/ — full package with Cargo.toml, pyproject.toml, src/lib.rs, Python package, stubs, tests
• Compiled with deepcut feature (default); embeds model/deepcut.onnx via include_bytes!
• Declares nlpo3~=2.0 runtime dependency
• DeepcutTokenizer accessible directly or via base shim — same call site, same results

CI

• build-python-wheels.yml: adds parallel build_wheels_deepcut / build_sdist_deepcut jobs; publish_pypi uploads both packages
• test-nlpo3-python.yml: adds test-python-deepcut job; base job no longer needs deepcut available

Docs

• docs/design.md: "Proposed split packaging" → implemented; updated with actual layout and behavior
• docs/implementation.md: new "Split Python packaging" section documenting shim mechanics and module names
• CHANGELOG.md: breaking-change note with migration instructions