This adds support for exact likelihood calculation, correct posterior
sampling of models with nonempty groups, much improved "distributed"
degree-corrected prior, as well as continuous edge covariates.

Now PropertyMap.get_array() will return a simple unchecked numpy.ndarray
subclass (PropertyArray) that will be invalid if the shape of the graph
changes. The user is required not to store a copy of this value if the
graph is to be modified.

This fixes issue #305

This fixes issue #300

Make sure to preserve "ignore_degrees" even when the values are
non-Boolean (i.e. 2).

This fixes issue #296.

Now algorithms rely more on parallel_{vertex/edge}_loop() and global
the paralellization threshold is honored.

This fixes issue #191