Investigate use of Array API

[matthewfeickert]

Description

In discussions with @rgommers, @jrbourbeau, @jakirkham at SciPy 2019 it was brought up that given the adoption of the (NumPy) Array API across multiple libraries that we should see how this can be used in pyhf, particularly with respect to CuPy (Issue #238) and Dask (Issue #259).

They would be interested in getting feedback on how useful this API adoption is, so we should also do a good job of noting this in our PRs and then given them a short writeup about our experience (or maybe a blog post).

[matthewfeickert]

As I'm the one proposing this work, I am happy to be the one documenting this it. Though we will need to address the other Issues first.

[lukasheinrich]

if something like this exists I'd be very excited to adopt this. The slight mismatches in APIs between the backends become quite annoying. Originally I was thinking if Tensorly for this #131.

[matthewfeickert]

The slight mismatches in APIs between the backends become quite annoying.

This won't help with the TensorFlow and PyTorch backends, as they have decided to not conform to the NumPy API and I don't see them radically changing their APIs. But this would help with everything else.

Originally I was thinking if Tensorly for this #131.

Yeah, I was looking at Tensorly the other day as well. This might be something to revisit in the future, but I'm also not sure if the API advantages that Tensorly would give would be worth taking it on as another core dependency that we would be very fragile to.

[kratsg]

Tensorly looks to be doing something similar to what we've been doing -- which is interesting.