Pzflow

Improved the stability of checking for flags during posterior marginalization. Now robust checks for floats and np.nan

Changes made:

- no longer automatically adding .pkl file extension when saving flows and ensembles
- fixed FlowEnsemble so that metadata (e.g. data_columns) is properly stored at the ensemble level
- changed nsamples to err_samples for error convolution in order to increase clarity
- added ability to convolve non-Gaussian errors
- officially added support for modeling discrete variables via UniformDequantizer
- added ShiftBounds bijector so I don't have to keep abusing StandardScaler
- added flag so you can turn off the automatic standard scaling of conditions passed to a conditional flow
- added get_ to the front of the example functions so that their use is more clear
- added support for column marginalization in posterior calculation
- expanded and reorganized the example notebooks

Fixed a bug in sampling from a conditional FlowEnsemble

Updates:
- defaulted B=5 in splines
- Sampling from error distribution for error convolution now available for training, log_prob, posterior
- Improved the way flows are loaded from a file to make more backwards compatible
- Added the centered beta distribution (CentBeta)
- Conditions are now automatically standard scaled before being fed to the neural networks inside the normalizing flow. The appropriate means and stds are calculated and stored during training.
- Added a FlowEnsemble class for easily training a deep ensemble of models
- Added example notebooks for error convolution and FlowEnsemble


v.1.6.0
Added an option for a random seed when initializing the flow. This allows you to create deep ensembles.

Moved flow saving to recursive dill pickling, so that more complex objects can be stored with flows.

Fixed the analytic error convolution for Gaussian base distributions.