Array class

[alewis]

It seems there is widespread agreement that, in the longer term, Node is eventually to be replaced by or supplemented with some sort of "I can't believe it's not an array!" class Tensor that is less stateful. Users would then by default interact with Tensor rather than with Node directly.

In the shorter term, to motivate adoption by existing physicist practitioners who will primarily interact with the library via ncon, we need something now that behaves more or less like an array, but which nevertheless shields the user from the backend. Originally, I had thought this might be done by modifying the interface to Node to include a subset of NumPy functions, but I now see this might be complicated, due to the need to account for the various possible Edge states in order to do e.g. reshape or transpose.

I therefore propose that we instead directly provide an array class (I suggest we call it Array), which in essence would just be a wrapper around the backend array. Node.tensor would then be an instance of Array instead of the backend array directly.

We provide a subset of NumPy-like functions to act on Array. Contractions can be handled using either ncon or by building Nodes and Edges around Arrays; users who only want to use Node would just instantiate them from backend arrays. I think this would simplify e.g. PEPS a lot; it would also lay much of the groundwork for Tensor, which would differ from Array at the interface level more or less only via the contraction API.

Thoughts?