Ensure ufunc/function dispatching is narrow

[seberg]

To allow potential wrapping e.g. via projects like pint, cuNumeric should only accept types in the __array_function__ and __array_ufunc__ protocols that it clearly understands. Right now, these are superclasses (if there was subclassing in theory) and NumPy arrays. CuPy arrays would also fall into this category.

Ping @manopapad since we talked about it briefly yesterday, thought I would just look into it. It is correct that cuNumeric currently only knows about NumPy arrays really (and objects that don't have __array_function__/__array_ufunc__); i.e. no cupy?

I would lean towards it being OK to just fail if you don't implement this. Dask has a FutureWarning for their NumPy fallback path in __array_function__, but it looks a bit like this may be important for other fallbacks and I think its fine.

[rapids-bot]

Pull requests from external contributors require approval from a nv-legate organization member with write permissions or greater before CI can begin.

[manopapad]

LGTM, giving a chance to @bryevdv to comment if he wants

[manopapad]

It is correct that cuNumeric currently only knows about NumPy arrays really (and objects that don't have __array_function__/__array_ufunc__); i.e. no cupy?

Right, cuNumeric can attach directly to the buffers of NumPy arrays. Today that doesn't mean much, since we'll likely have to make a copy right away to move the data to the GPU, and/or partition the array for the needs of the next operation, but the ability is theoretically there to use the memory directly. And in any case, the updates will be automatically mirrored on the original NumPy array.

This also simplifies how we handle non-array values where an array would be expected. In that case we just call np.array/np.asarray on it, then attach to the result. This way we automatically handle things like scalars and lists.

IIRC cupy doesn't like it when np.array is called directly on its arrays, because that is a silent conversion with heavy performance implications (they would rather users do a more explicit conversion). Similar to NumPy, we could probably attach to cupy arrays directly (by taking advantage of the CUDA Array Interface), but we're not doing that today.

[bryevdv]

LGTM