re-arange the code with a _svi_model abstract class superseeding all the other [model]Flow class + start implementing PLNPCA flow

Implement Sylvester NF + change encoder dim and try glorot init of Network weights : still gives inf.

1. refactor code in submodules for _amortization.py (encoder, variationaldist), _flows.py & 2. start introducing a optional non standard Gaussian distribution for the NF & 3. Use a common encoding h(x) in dimension 50: all neural net work from this common h(x).

simulation setting with (2, 2) as offset, no covariate. PlnSto works decently bu still don't manage to make NF work

sampling.py: put a setter for the  property in PlnParams class. Useful to modify a simulation setting.