benchmark: establish systematic linear and GLM inference coverage

[TheHiddenObserver]

Goal

Add methodical benchmark and validation coverage for maintained linear-model and GLM inference outputs.

Scope

• align coefficient, covariance, standard-error, test-statistic, p-value, confidence-interval, and likelihood criteria with appropriate external references;
• cover NumPy/CuPy/Torch paths and strict failure behavior;
• separate inference correctness from fit timing;
• retain model, family, covariance mode, scale, tolerance, and seed identity;
• produce structured sources, parser tests, registration, generated assets, and coverage-matrix updates.

Parent: #100