Implement a GumbelSoftmaxReparam

[fritzo]

This adds a reparametrizer for the Gumbel-Softmax trick.

The reparametrizer replaces a d-dimensional RelaxedOneHotCategorical with a d-dimensional Uniform(0,1) distribution whose samples are first transformed to be Gumbel distributed and then passed through the softmax trick.

The motivation is that, when unimodal autoguides try to learn the posterior of a R.O.H.C. distribution they are only able to capture a single mode, i.e. a single category. By instead learning the posterior in pre-softmax space, a single mode can then be softmax-transformed to a multimodal posterior, with one mode per category. Thus this reparametrizer allows naive autoguides like AutoNormal to capture multimodal posteriors.

Tested

• added moment tests
• smoke tested in a practical model

[fritzo]

Thanks for reviewing!