Label handling commit breaks the imdb finetuning script

Thomas, thanks for sharing this code! I noticed that commit 8d9c2371fc37ba8958f174501a9afb91a7ef7a06 seems to have broken the default functioning of the classification finetuning scripts - in the previous version there seems to have been a key called `'labels'` associated with the imdb and trec dictionaries, but in `finetuning_train.py` [this line](https://github.com/huggingface/naacl_transfer_learning_tutorial/blob/69031641946979f0ce1a0961a6b50e290f3de434/utils.py#L127) still references the now deleted key.

I updated the line to just use `DATASETS_LABELS_URL['imdb']['test']` as intended, but then it seems that the S3 bucket doesn't have the IMDB test file.

See below:
```
file_path = "https://s3.amazonaws.com/datasets.huggingface.co/imdb/test.labels.txt"
label_file = cached_path(file_path)
with open(label_file, "r", encoding="utf-8") as f:
all_lines = f.readlines()
print(all_lines[:5])
```
Gives:
```
['\n', 'NoSuchKeyThe specified key does not exist.imdb/test.labels.txt3D9E7C511167A0FBRiidOcrHfFaqxW9tmUXRppE/G3lsYoCZcq+uaYDi2yPPoe8mv/Og6PMuUncwk+B53tGsvcCZMWk=']
```

Does the test file for IMDB still exist with this name? This doesn't seem to be an issue with TREC.

Start in finetuning_train.py and the DATASETS_LABELS_URL definition referenced around utils.py line 127, then run the IMDB classification finetuning path. Check whether the configured S3 path for imdb/test.labels.txt matches an available test file and whether the default script completes without the deleted labels key. Done means IMDB and TREC label handling both work with the current dataset metadata.

Rédigé par le modèle d'indexation à partir du texte de l'issue.