What is the motivation for this task?

Export is not working for the WinCLIP implementation because the masks and reference embeddings are currently not stored in the model's state dict.

Describe the solution you'd like

Enable export for WinCLIP model by registering masks and reference embeddings as buffers so that they end up in the state dict.

Additional context

Note that storing the masks and reference embeddings in the state dict is not necessary for testing and predicting, since we re-collect these attributes at the beginning of the test and predict sequence in the setup call of the lightning model.