fix #756 

Some models do not need inputs during inference, e.g. GANs. However,
it was necessary to feed dummy-inputs although they are not needed.

Do not trust trainable_variables create by Keras layers. Use `M.weights` instead and print warnings about unknown variables. (#748)

Doing:

```
ds = RepeatedData(ds, -1)
ds = PrefetchDataZMQ(ds, num_workers)
ds.reset_state()
```

Would blow up, since `RepeatedData` raises a ValueError, and `PrefetchDataZMQ` only checks for `NotImplementedError`.

Not the cleanest fix but 🤷‍♂️

* Removed excess copying when fetching data from ZMQ

The `.bytes` explicitly triggers a copy in order to create a Python `bytes` object. This isn't necessary for pyarrow deserialization.

* Remove all the `.bytes`!

* Fixed crash when deleting a MultiThreadPrefetchData instance when `reset_state()` was not called