It looks like something is trying to deepcopy an object that includes a Python type Union, which is a known Python issue. Not sure where that happens and where the type would come from, though
In any case, there's not really a point in using a transformer starter model for ner.teach. The entity recognizer you train won't use the transformer weights anyways (you're just training a regular spaCy NER model) – and even if it did, updating/finetuning the model in the loop would be way too slow and not really effective, as the transformer models typically need larger batch sizes and need more than just a simple, single update with a small batch. So the model wouldn't really change as you update.