Generalized deconv

[dehann]

resolves discussion on #464 as well as fully non-Gasussian operation.

Each variable type---i.e. domain / manifold---can also belong to a variety of (mathematical) Groups used to define factors. For example range bearing defines a group between Pose2, Point2 in local or Point2, Point2 in global reference frame.

Deconv should automatically select (using a "Group Tree") the best kinematics to define the transformations on the group.

[dehann]

also resolves #577

[dehann]

xref #459

EDIT: using LikelihoodMessage to extract which probabilistic chain rule expansion of :UPWARD_DIFFERENTIALs to describe a message joint internal structure.

[dehann]

required to resolve #501

[dehann]

This only works for two variable relative likelihoods at this point, so further generalization of is required. Current status is far enough to close this issue.