Invertible Gaussian Reparameterization and Logistic-Normal approximation of Dirichlet

[kisnikser]

This PR focuses on implementing Invertible Gaussian Reparameterization trick with corresponding distribution, and custom Logistic-Normal distribution with two-side approximation with Dirichlet distribution.

Some details about closed-form Laplace bridge for Logistic-Normal and Dirichlet distributions

• We implement a LogisticNormalSoftmax distribution, which is a transformed distribution from the Normal one. In contrast to original LogisticNormal from pyro or torch, this one uses SoftmaxTransform, instead of StickBreakingTransform that allows us to remain in the same dimensionality.

• We implement two distinct functions, each of them has a distribution on input (Dirichlet or Logistic-Normal), and returns an approximation distribution.