Refactored code and created tensorflow backend

[chaserileyroberts]

This is a big change, but I think it is for the better. (See #51 for details).

The next backend we will add will be a pure numpy implementation, then JAX, and then leave torch to the community.

[chaserileyroberts]

This is a big design change so I would appreciate your inputs.

[sleichen]

network_components and ncon_interface need to be updated as well, right?

Also, I'm not sure how switching backends will be implemented, but it seems conceivable that someone could try to merge two networks created using two different backends. How do we deal with that?

[sleichen]

trace --> outer_product

[chaserileyroberts]

Can you tell I copy and pasted? lol

[chaserileyroberts]

Also, I'm not sure how switching backends will be implemented.

Two ways (I'm waiting on stavros building the numpy backend before trying it).

1. Doing tensornetwork.TensorNetwork(backend='numpy')
2. Defaulting to the global backend, and changing the global backend with tensornetwork.set_backend('numpy')

but it seems conceivable that someone could try to merge two networks created using two different backends. How do we deal with that?

Easiest thing to do is throw an error. We can check that the backends are the same before merging, which is pretty easy. We could have a converter by changing each node's tensor by doing torch.Tensor -> np.array -> tf.Tensor, but I expect that would get messy and I'm not sure people would use it often enought to justify the complexity.

All 4 backends we plan on supporting right now (JAX/PyTorch/TF/numpy) allow tensors to be created by converting a np.array to their own Tensor type. All higher level network creators we build would just have to build nodes using numpy and they would get converted internally.

[chaserileyroberts]

network_components and ncon_interface need to be updated as well, right?

Yes they do.

[sleichen]

LGTM