Workaround mirror for slow model downloading

The `roberta-base-nli-stsb-mean-tokens` was taking 2+hrs to download for me so i uploaded it to huggingface/transformers generously open and 100x faster s3 bucket if this is useful for anyone.

Usage:

```
from sentence_transformers import SentenceTransformer
from sentence_transformers.models import RoBERTa
from sentence_transformers.models import Pooling

roberta = RoBERTa('pertschuk/0_RoBERTa')
pooling = Pooling(768)
model = SentenceTransformer(modules=[roberta, pooling])
```

No repository file, test, or entry point is named. Start by determining whether this workaround is intended as a project change or only a usage suggestion, then review the shown SentenceTransformer and RoBERTa usage with maintainers. Done requires an explicitly defined integration or documentation outcome.

Written by the indexing model from the issue text.