Mark DebiasConvex non-negative modes as experimental and disable inference outputs

[fedemolina]

Summary

• Mark DebiasConvex(method_non_neg="svd") as an experimental point-estimation heuristic.
• Return std=None, inference_valid=False, non_negative_method="svd", and diagnostic fields whenever the non-negative heuristic is requested.
• Reject unsupported method_non_neg values, including nnmf, with a clear top-level error because that path changes the model family and is not currently compatible with DebiasConvex debiasing/inference.
• Keep raw DebiasConvex mode backward compatible with standard errors and inference_valid=True.
• Add focused tests, a slim local documentation page, and an executable tutorial notebook.

Notes

This PR does not claim inference-valid non-negative DebiasConvex estimation. It adds guardrails so the experimental non-negative point-estimation path is explicit, diagnosable, and does not report standard errors as valid.

Related context: #12

Validation

• env PYTHONPATH=src poetry run pytest tests/test_debiasconvex_nonnegative_guardrails.py -q
• env PYTHONPATH=src poetry run pytest src/causaltensor/tests/test_synthetic_class.py::TestSyntheticClass::test_dcpr src/causaltensor/tests/test_synthetic_class.py::TestSyntheticClass::test_dcpr_multiple src/causaltensor/tests/test_real_class.py::TestRealClass::test_dcpr -q
• env PYTHONPATH=src poetry run pytest tests/test_debiasconvex_nonnegative_guardrails.py src/causaltensor/tests -q
• poetry run python -m py_compile src/causaltensor/cauest/DebiasConvex.py tests/test_debiasconvex_nonnegative_guardrails.py
• Executed tutorials/debiasconvex_nonnegative_guardrails.ipynb from repo root and from tutorials/ cwd.