Array API support

[tomwhite]

It would be good to allow the arrays in IntegerArray and BooleanArray to be any Array API-compatible array, rather than only NumPy arrays.

(The motivation here is that in Cubed we use ndindex to canonicalise indexes and need to convert to and from NumPy arrays for integer array indexing, since the underlying array could be any that supports the Array API.)

[asmeurer]

This is a good idea. I think the most difficult thing is coming up with a way to hash the array. We use hash(array.tobytes()) for numpy (and I'm wondering if there isn't a faster way than that).

ndindex/ndindex/array.py

Lines 168 to 169 in 5c00707

	def __hash__(self):
	return hash(self.array.tobytes())

For the array API there isn't a native way to do this. Maybe something can be done with dlpack. We do also use Cython now so if something needs to be written in C that's an option. Although only strided array backends will support dlpack, so this isn't a completely perfect solution (but then, ndindex is only going to work for completely realized, non-lazy arrays so maybe this is fine).

We also set the numpy array to be read-only, but that isn't really a requirement. We always copy the array on input so that should really be enough. As far as the operations done on the arrays themselves, barely anything is used and the array API should be more than sufficient https://github.com/search?q=repo%3AQuansight-Labs%2Fndindex%20.array&type=code.

[asmeurer]

We could also try something like https://stackoverflow.com/questions/74805446/how-to-hash-a-pytorch-tensor