Factorized log probability method, general variational bound class

[suhasshrinivasan]

We want the log probabilities of a random variable dimension by dimension when the distribution is factorized for a few cases such as density inspection. This PR adds factorized log probability computation methods in the core classes such as IndependentTrainableDistributionAdapter and also extends the idea to factorized log dets in transforms in normalizing flows and factorized elbo and iw bound in variational dequantized distributions. The functionality is lightly tested and elucidated in a jupyter notebook in notebooks/factorized_log_prob.ipynb.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[suhasshrinivasan]

I decided to rollback ELU implementation, since changing its constructor broke backwards compatibility. We would need to discuss the modification once again.

[suhasshrinivasan]

In gensn/parameters.py, numerical underflow was avoided by adding a generic epsilon value of 1e-16 and this breaks at times with PyTorch's operations (I faced a numerical underflow with eigenvalue computation). It is now changed to adding a dtype specific eps (e.g. torch.finfo(torch.float32).eps) which resolves the issue and is generally a more standardized solution.