Make init_par a class Method

[ilan-gold]

Right now, initialization of the parameters occurs on a per-model basis but generally the function is quite similar across different models. This should become a class method.

[picciama]

I don't necessarily agree, because it doesn't only depend on the model but also on the specific "flavour". edgeR for example uses some fancy QR decomposition stuff to initialise the location params, while the dispersion params are fixed to e.g. 0.05, which needs to be translated to our case, which is then np.log(1 / 0.05) over all features. Let's wait till the stuff from #145, is merged.