Experiment with lazy imports instead of local imports in tensor-network backends

[garrison]

The tensor network backends deliberately avoid importing the actual backends, because they are slow to load (especially quimb), and we do not want to pay that price until they are used. Currently we deal with this using local imports in each function. I'd be nice to explore the possibility of making them be lazy imports (via wrapt) instead.