Numerical variations in IRLS for Poissons

[matteodelucchi]

irls_poisson_fast.cpp results in slightly different score values compared to the same model computed with glm.

Steps to Reproduce

Compare mycache.mle with modglm in test-build_score_cache_mle.R .

Current Bug Behaviour

> mycache.mle$mlik
[1] -1418.438      -Inf
> logLik(modglm)
'log Lik.' -1410.645 (df=2)

Analogous for AIC and BIC scores.

Actual expected Behaviour

I'm unsure if this variation in score values is expected.

Relevant Logs

This was temporarily fixed with an increased tolerance to pass the tests.

Possible Solutions

Double-check IRLS Poisson Fast algorithm. It has been shown that numerical overflow is not handled properly for large values of eta. Unsure if eta should ever be that large or if this was only caused by a faulty test. If the latter, consider catching such cases upstream properly and investigate why glm did not raise a warning.