get preds off gpu to improve perf (#269)

We were leaving the label predictions tensors on the GPU,
which was leading to lots of expensive GPU calls to read
the data. In particular, `tt profile` revealed the following:

![Screen Shot 2023-07-19 at 11 41 06 AM](https://github.com/allenai/mmda/assets/1287054/9f6736d2-51be-4d65-8bfb-fe93cca9eed7)

Scalene's GPU time reporting is generally full of false attribution,
but it was a hint in the right direction in this case.
Based on my interpretation of the code, we were being forced
to access this tensor off the GPU three times for _every_ input
word in the document.

After pulling the label preds into system memory:

![Screen Shot 2023-07-19 at 11 44 24 AM](https://github.com/allenai/mmda/assets/1287054/5853924a-824b-4818-aeba-fe121d48c9af)

I confirmed the remaining high GPU items reported by the
profiler have nothing to do with the GPU, 
and the code is too scary to futz about with anyway.

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = 'mmda'
-version = '0.9.5'
+version = '0.9.6'
 description = 'MMDA - multimodal document analysis'
 authors = [
     {name = 'Allen Institute for Artificial Intelligence', email = '[email protected]'},

src/mmda/predictors/hf_predictors/mention_predictor.py

@@ -107,7 +107,7 @@ def predict_page(self, page: Annotation, counter: Iterator[int], print_warnings:
                 )
                 batch.to(self.model.device)
                 batch_outputs = self.model(**batch)
-                batch_prediction_label_ids = torch.argmax(batch_outputs.logits, dim=-1)[0]
+                batch_prediction_label_ids = torch.argmax(batch_outputs.logits, dim=-1).tolist()[0]
                 prediction_label_ids.append(batch_prediction_label_ids)
 
         def has_label_id(lbls: List[int], want_label_id: int) -> bool: