Support keras/pytorch/other model serialisation in kubeflow backend

There are some issues with serialising keras models using `dill` - in particular, deserialising a Lambda layer seems to bomb. Should also check that we can serialise pytorch models, given that they're becoming a standard.