Description

This PR implements an optimization for the GLM Hessian calculation (Observed Information Matrix).
When a canonical link is used, the Observed Information Matrix (OIM) coincides with the Expected Information Matrix (EIM) / Fisher Information.
Mathematically, the curvature term involving$(y - \mu)$ vanishes. This PR detects canonical links and skips this unnecessary computation.

Closes#4269

Changes

1. Family Logic: Addedis_canonical_link property toFamily class and_canonical_link_class to subclasses (Binomial, Poisson, Gaussian, etc.).
   • Note: I usedtype(link) is ... instead ofisinstance intentionally to strictly match the canonical link implementation and avoid inheritance issues.
2. GLM Optimization: Updatedhessian_factor ingeneralized_linear_model.py. Ifis_canonical_link is True, it setsobserved=False, forcing the use of the EIM formula ($-X^T W X$), which is faster and numerically equivalent.
3. Documentation: Added detailed NumPy-style docstrings with mathematical justification inhessian_factor.
4. Testing: Addedstatsmodels/genmod/tests/test_glm_canonical.py with 11 tests covering all major families to ensure numerical equivalence between the optimized path and the original path.

Verification

• Ran new tests:pytest statsmodels/genmod/tests/test_glm_canonical.py (11 passed).
• Verified code formatting (flake8).
• Docstrings added with LaTeX math support.