Add an option to require a GPU

[mdekstrand]

We should add (or document) an option to fail instead of slowing down when a GPU is not available.

• Verify that LK_DEVICE=cuda causes training to fail when no CUDA device is available
• Support setting the device in lenskit.local.toml
• Optional: support .env in LensKit, possibly behind an option in lenskit.toml
• Document setting LK_DEVICE
• Add warnings when training without CUDA for models that are very slow on CPU