Estimator abstractions

[manuelgloeckler]

We currently only have the DensityEstimator abstraction, but we might want to expand this for other approaches. I would suggest something like this:

class Estimator(nn.Module):
    def loss(...)

class DensityEstimator(Estimator)
    def sample(...)
    def log_prob(...)
    def sample_and_log_prob(...)

class RatioEstimator(Estimator)
    def log_ratio(...)    ( Not sure, depends on what we need, here)

class ScoreEstimator(Estimator)
    def score(...)

The different Posterior Classes will implement the sample and log_prob methods anyway and must be adapted to support different estimators internally (if possible).