Prevent tok2vec to broadcast to listeners when predicting

[svlandeg]

Description

Addressing the same problem as laid out in explosion/spacy-transformers#342, i.e. the fact that tok2vec.predict broadcasts the outputs to the listeners, breaks the usual contract in which the training mechanism should rely on that broadcasting, but in prediction we should use doc.tensor.

Because we still want to be able to have a frozen (but annotating!) embedding layer, have a fallback in the listener's forward function that will access the doc.tensor even when training, if there is otherwise no valid data available.

Types of change

bug fix

Checklist

• I confirm that I have the right to submit this contribution under the project's MIT license.
• I ran the tests, and all new and existing tests passed.
• My changes don't require a change to the documentation, or if they do, I've added all required information.

[svlandeg]

We might want to move this to develop, but I wanted to open this PR first to instigate some internal discussion.