Improve Taylor models and similar structures

[reverendbedford]

We need a way to "wrap" a batch dimension in some reduction/expansion operations. So a model of this type might expand the base model forces along some axis, run the batch, but only return the average result. The external interface to this type of model is consistent -- you have the same batch shapes in and out. The only real challenge is the "inner" model state, which will have one extra batch dimension and, for stateful models, need to persist from time step to time step.

Solving issue #116 may help with this.