Find better Poisson approximation in ML backends

[matthewfeickert]

Description

As I forgot in Issue #266, the different backends have different approximations of the the Poisson distribution. In the NumPy backend a Gamma function based approximation is used. However, in the TensorFlow, PyTorch, and MXNet backends a Normal function approximation is used.

It would be good if possible to try and have a better approximation than the Normal approximation that is more consistent across backends. I forgot this, and so it won't be obvious to new users.

[lukasheinrich]

many backends support pdf distributions. The only possile issue is that poisson is nominally a discrete pdf. but iirc, most backends should already use a continuous approximation of it anyways.