Refactor `TensorNetwork` type, and add a reverse index mapping.

[jack-dunham]

The new TensorNetwork type should

• Have its edges strictly defined by the indices of its tensors.
• Not necessarily rely on an underlying_graph.
• Accept a dictionary of tensors, with keys as the vertices, as its primary method of construction.
• Have a index (tensor dimension) to vertex mapping.

As a corollary to the above, a TensorNetwork should

• Not implement rem_edge! (see this comment)