Adapt SNRE to new DensityEstimator abstraction

[gmoss13]

It would be good to create a RatioEstimator to use for SNRE methods, similar to DensityEstimator abstraction already implemented. The API and broadcasting should be similar, so possibly this should be handled in a parent Estimator class (see #966)

[bkmi]

I will take a look at this and formulate some ideas.

[bkmi]

The purpose of the layer is to:

1. compute the loss
2. wrap the ratio estimate (for example for weighing it)
3. embedding functions for both theta and x (subsuming StandardizeInputs #1036)
4. provide a function to combine embed_theta(theta) and embed_x(x) into a single tensor.

Non-goals:

• introduce methods to approximate ratios of normalized distributions as in torch-mist here

Progress:

• [x] Change classifier_nn helper function to return new Ratio Estimator Class
• [x] Integrate new interface into SNRE base classes.

[bkmi]

Note: this will address the ratio aspect of #957 .

[bkmi]

this is being updated to reflect the goals of #1041 and #1046

[janfb]

solved with #1097