Poisson Model

[ilan-gold]

It appears that for the test that is failing, the threshold is .2 for the mean and the real average deviation is 0.27 - I've heard previously that these tests can be unreliable though so not sure how to proceed. A few other things:

1. I largely copied the closed initializers (since I still don't fully understand what they are doing) and removed the scale aspect of them
2. I am not sure how the bounds method is derived so I copied that as well - is there something in the manuscript?
3. Similar to the normal case, we have arguments we don't need - here though, we don't need anything related to scale. I think given that 2/3 of the models that we use now deviate from the standard implementation, maybe all of this should be in kwargs? Not sure what the idiomatic python thing is to do.
4. If the failing test is indeed wrong, what are some good ways to debug here?

[ilan-gold]

Tests fixed!

[picciama]

Overall I'd say this is very much done. Only some minor things I've spotted.
Most importantly, I'd say init_theta_scale should really be removed and maybe we can reuse some functions by putting them into the base_glm submodule. But that might be better to do once all the noise models are implemented and merged.