Optional autodiff support?

[alxmrs]

It would be awesome if the backing array implementation supported auto differentiation, that we could access some grad method from Cubed.

It looks like a bunch of stakeholder libraries have this functionality:

https://data-apis.org/array-api/latest/purpose_and_scope.html#stakeholders

Though, differentiable programming may be out of scope for Cubed. @TomNicholas @tomwhite @rbavery any thoughts here?

I have a pipe dream of turning Cubed into an ML framework, and I think this would play an important part.

I haven’t thought of all the implications, but a potential sharp edge that @shoyer once pointed out to me: there will probably be significant memory differences between an op graph and its gradient. Can Cubed’s spec model be extended here?