default_device and default_dtype fail with NumPy < 2.1 #1078

Based on CI failures I've been getting in scipy/scipy#26550, I believe these will fail:

import array_api_extra as xpx
import numpy as np
xpx.default_dtype(np)  # AttributeError: module 'numpy' has no attribute '__array_namespace_info__'
xpx.default_device(np)  # presumably would fail, too

I'm having trouble testing in Colab because I get an older array-api-extra with NumPy 2.0. Anyway, it makes sense, since NumPy only added __array_namespace_info__ in SciPy 2.1.

On a related note, it would be nice to have a function that returned that array-api-compat version of a namespace given the namespace. One can do, e.g., array_namespace(xp.asarray(0)) # skip device check but creating temporary arrays is exactly what default_dtype and default_device saves us from. Perhaps array_api_compat.array_namespace should also just accept backends?