Add possibility to use opt_einsum instead of pure Numpy

[dtellenbach]

One of the most challenging parts of contracting a tensor network is to find a good contraction ordering. Currently ncon just uses some simple default ordering or a user provided one. opt_einsum is a highly optimized tensor network contractor that is capable of finding very good contraction sequences.

To be able to use the ncon interface with a high quality network contractor, this patch adds two new arguments to ncon:

1. backend, a string that can be set to numpy or opt_einsum to enable using opt_einsum. Defaults to numpy
2. opt_einsum_strategy, a string to select a contraction strategy for opt_einsum. Defaults to auto.

[dtellenbach]

Still a draft since I haven't updated the README yet.

[mhauru]

Hi @tellenbach! What's the advantage of opt_einsum over Numpy's own einsum with contraction order optimisation?

[dtellenbach]

Hi @mhauru! Although many of opt_einsum's strategies to find good contraction sequences have been upstreamed into Numpy's einsum it still has newer and more strategies.

Another big advantage of opt_einsum over einsum is that the contractors provided by opt_einsum are ready to be automatically differentiated using e.g. autograd or Jax. This is e.g. interesting in the context of optimizations of tensor networks on Riemannian manifolds.