fix: ak.from_numpy should fail on zero-dimensional arrays.

[tcawlfield]

At this point, the first tests, ak.Array(), was already passing without modifications.
But the second test was not -- ak.from_numpy().
Fix is in _layout.py. No policy options are passed to _layout.from_arraylib().

[tcawlfield]

Ugh this was too naive. Breaks a bunch of other tests.

[tcawlfield]

Ugh this was too naive. Breaks a bunch of other tests.

Okay, I fixed the problems by adding a primitive_policy kwarg to _layout.from_arraylib and upstream functions.
I'm not sure now why all the CI tests are failing. They pass on my machine, and the reasons seem to relate to C++ issues outside the scope of my change here.

This change ended up being noisier than I would have liked, but I don't see any other way to get the desired policy.

[jpivarski]

If from_arraylib were a public function, I'd want to think about the consequences of adding a primitive_policy argument. As an internal function, we can always reconsider it's interface at a later date, and this interface looks good enough for now. But the defaults seem to be inconsistent, and that would make it hard to think about.

Understanding the test failures will be important, since they're telling us that behavior is changing—probably not in ways that we want.