add predict_step to BYOLTask

[isaaccorley]

Related to #813. This PR adds a predict_step method to BYOLTask so that users can utilize PyTorch Lightning's trainer.predict() function which will automatically loop and predict over a PyTorch DataLoader e.g.:

preds = trainer.predict(model=task, dataloaders=datamodule.test_dataloader())

# Or if your datamodule has a `predict_dataloader` method defined
preds = trainer.predict(model=task, datamodule=datamodule)

Note the way the EncoderWrapper is setup is that calling self(x) on a batch of images will return the projector output embeddings. However, we only need this during training. For linear evaluation or other downstream tasks we want the output of the image encoder (the output of the avg pool layer of the ResNet). To work around this I had to add another hook to store the embeddings on a forward pass in the self._embeddings attribute.