Motivation: as we expand the scope of models supported, it may be easier for contributors to add new distributions if we refactor the class to accept distribution-specific params and corresponding likelihood / loss / gradients / hessians as callables.

One strategy inspired by statsmodels proposed by @jasmainak in #274 is as follows:

distr = NegativeBinomial(n_failure)
GLM(distr=distr)

class MyDistribution(BaseDistribution):
    def loglikelihood(self):
        pass
    def loglik_gradient(self):
       pass
   def loglik_hessian(self):
       pass

There are also alternatives such as the scikit-learn approach which is less work for the low touch user who only needs to call fit and predict on a model object without having to construct a distr object:

model = PoissonGLM(base_param1, base_param2, poisson_param1, poisson_param2)

class PoissonGLM(GLM):
    def loglikelihood(self):
        pass
    def loglik_gradient(self):
       pass
   def loglik_hessian(self):
       pass