Handle possible NULL pointers in the array buffer

[ngoldbaum]

The discussion around numpy/numpy#23262 made me realize that there can be NULL pointers in the array buffer after np.zeros and np.empty. This means that we need to check for that case whenever we access an existing string array buffer in the casts and ufuncs.

To support this I added a new helper empty_if_null, which returns a new empty static string. I couldn't figure out how to do this in a way that avoids heap allocations because the ss struct uses a flexible array member, so I also need to free the resulting empty string later.

It would be possible to avoid this if there was a way for a dtype to supply an initial value for np.empty and np.zeros.

Alternative suggestions for how to handle this are very welcome.

[seberg]

So, if we definitely don't do reference counting (each array that is not a view owns a copy of the string), then I do still think it makes sense to move the length information into the array data itself.

On this case, how about a pattern off:

    const ss *s = load_string(in);

And then return a statically allocated empty ss (or a once allocated global). That way you don't need the free() call. Freeing is only needed on store, and there you can do it unconditionally. The only thing is you must not free what get_string returns. I hope the const will protect a bit and it should at least be a nice hard crash that is easy to track down.

[ngoldbaum]

Closing this in favor of a new PR implementing Sebastian's suggestion.