Launching with an unsupported GPU

I'm using an Nvidia GTX 1080 which only supports up to CUDA 6.1. PyTorch wants at least CUDA 7.0.

I applied a forced device override to line 129 in ``/home/mark/.embedding-atlas/lib/python3.12/site-packages/embedding_atlas/projection.py`` and I was able to train and launch without issue.

```python
transformer = SentenceTransformer(model, trust_remote_code=trust_remote_code, device='cpu')
```

It might be worth adding a ``--device`` CLI flag for anyone else wanting to try this package out without needing a CUDA 7+-supported GPU.

Start at projection.py around line 129, where SentenceTransformer is created with a forced CPU device, and trace how launch options reach that code. Add a way for the launcher to accept a device choice, then verify that CPU launching works on hardware without a supported CUDA version.

Written by the indexing model from the issue text.