Bug: numpy default rng persistence

[BenjaminBossan]

EDIT: I see what's going on, I'm on it.

When trying to determine the unsafe set of a numpy random Generator, there is an error. To reproduce:

def test_default_rng():
    import skops.io as sio
    rng = np.random.default_rng()
    sio.loads(sio.dumps(rng), trusted=[])

We test this class but only with trusted=True, which is why we didn't encounter the error in our test suite.

I haven't dug into the source of the problem, but possibly it's related to Generator being a cython class.

Maybe we should tighten our tests by relying as little as possible on trusted=True.