Draft: Add support for conditional JointPriors

[JasperMartins]

This fairly big PR updates ConditionalPrior, ConditionalPriorDict, JointPrior, and MultivariateGaussianDist. Since the changes are highly intertwined, I opted for only one PR. The main goal of this PR is to make it possible to use ConditionalJointPriors.

The changes are the following:

1. A bug in ConditionalPriorDict.rescale is fixed. Previously, it would not have been possible to rescale a subset of keys. Also, the check if the setting of the reference_params as done for ConditionalPriorDict.sample supports array is applied for rescale as well. Further, both ConditonalPriorDict.rescale and sample were updated to no longer need to create a copy of the ConditionalPriorDict with the requested subset of keys.
2. ConditionalPriorDict.check_conditions_resolved now checks if JointPrior.dist dependencies are resolved as well. If a JointPrior is a dependency of another prior, all JointPriors associated with the dist of that JointPrior need to be evaluated before the dependency is evaluated - otherwise, rescaling would always fail.
3. ConditionalPrior is updated to support JointPriors. This required accepting the named argument dist and a bit more flexibility on which named arguments to pass towards the init-function of the superclass. This also resulted in a change to the method get_instantiation_dict, which now only collects properties required to initialize the superclass and not all (named) arguments of ConditionalPrior.
4. JointPrior is updated to be able to set properties of BaseJointPriorDist-subclasses if these have explicit setter methods for these properties. That way, it ensures that no properties are changed in ways that would potentially break BaseJointPriorDist-subclasses.
5. MultivariateGaussianDist is updated to allow setting the mode to sample/rescale/calculate. This makes it possible to use ConditionalPriorDict to have a prior select the current mode. Using multimodal multivariate Gaussians for sampling is more straightforward that way, as it allows modes to "die off" during sampling if they do not contribute to the posterior. The implementation also supports arrays of different modes for different samples. Further, a few changes are made to _rescale, _sample and _ln_prob to be more efficient for large numbers of samples.
6. A bug in BaseJointPriorDist is fixed where bounds were not applied correctly.
7. Test cases for the introduced functionality are added.

Depends on #863