Pyprob

* Parallel sampler for importance sampling
* Copying of concatenated empirical distributions

* `Model.condition` and `Empirical.condition` that return conditional distributions based on a binary criterion.

* Update `ppx`

* New feature: Adding arbitrary log-probabilities with `pyprob.factor`. See "3.2.1 Conditioning with Factors" in van de Meent, J.W., Paige, B., Yang, H. and Wood, F., 2018. An introduction to probabilistic programming. arXiv preprint arXiv:1809.10756.
* New feature: Support on-disk target for `thin`, `resample`, `filter`, `map`, `reweight`
* `pyprob.observe` returns the observed value
* Improve resampling of weighted on-disk distributions
* Improvements and bug fixes in `reobserve`
* Use `sqlite` instead of `dbm` for on-disk empirical distributions and datasets. This removes the gnudbm dependency for faster disk operations.

* New feature: `Model.filter` to express constrained models
* New feature: `Trace.variable_sizes` gives a list of variables in a trace, sorted by their memory usage
* New feature: `Empirical.reweight` allows recomputing the weights of a weighted Empirical
* New feature: `Empirical.reobserve` allows modifying the likelihood distributions of an already sampled weighted posterior distribution (from an importance-sampling-based inference engine), so that likelihood terms can be tuned/calibrated without re-running the model prior. Idea by Giacomo Acciarini.
* Raise an error if the observed value is None in posterior conditioning
* Print effective sample size on-the-fly while sampling posteriors with importance-sampling-based inference engines
* `Model.sample` returns a `Trace` object sampled from the model prior (equivalent to `Model.get_trace` which will be deprecated)
* Added Bernoulli support for inference compilation
* Removed the `replace` feature from the pyprob.sample API
* Exclude tagged variables from diagnostics
* Inference network layers are not pre-generated by default when training with OfflineDataset
* Added support for moving distribution, variables, traces between compute devices
* `OfflineDataset.save_sorted` supports moving dataset between devices

* Fixed bug in incorrect covariances in Empirical density estimation
* Fixed default resolution of probability density plots
* Finalize empty Empirical distributions from `.filter()` where the condition is never satisfied
* Allow tagging (`pyprob.tag`) of arbitrary Python types. Previously only numerical values that can be converted to PyTorch tensors were supported.
* Support membership operator for named variables in a trace, e.g., `if 'var_name' in trace`