add `array_namespace` option to return wrapped if possible, else native

[lucascolley]

Since array_namespace(x, use_compat=None) now returns plain numpy, but released NumPy still has incompatibilities with the standard, we want to use use_compat=True in SciPy. However, since array-api-strict isn't wrapped, array_namespace throws a ValueError, so we have to use:

try:
    return array_api_compat.array_namespace(*_arrays, use_compat=True)
except ValueError:
    return array_api_compat.array_namespace(*_arrays, use_compat=None)

array-api-strict isn't special here, this goes for any standard-compliant library. It would be nice to have an option to do this without having to catch an exception. Alternatively, we could just wait until NumPy is fully compatible with the standard before making the native namespace the default, but I'm not sure how feasible that is as the standard continues to change.

x-ref scipy/scipy#21264

[asmeurer]

Well I think we're going to have to get rid of the "numpy 2.0 doesn't get wrapped at all" logic because there actually are some differences from the standard in NumPy 2.0 and even NumPy 2.1 (which weren't caught for some reason #167). I thought it would simplify things to not wrap NumPy 2.0 at all, but I think instead we are going to just have a separate, simpler wrapper namespace for it.

[asmeurer]

If we do end up needing more control of this option we can update it, but for now I think this should solve your issue. #170