Try out integration with Tensor Computations

[saulshanabrook]

The Tensor Computations library provides a succinct and efficient way of writing computations on arrays. It also agnostic to the memory backend. I think we should be able to provide an xnd wrapper, so that memory is allocated in xnd, but can be computed with TC. See https://facebookresearch.github.io/TensorComprehensions/integrating_any_ml_framework.html

We have to be able to attach some metadata to xnd tensors to be able to wrap them in the dlpack format: https://github.com/dmlc/dlpack

[saulshanabrook]

Possibly relevant issue around nesting function computation facebookresearch/TensorComprehensions#56