Update the device doc string to include multi-gpu inference.

[oscarkey]

We could enable multi-gpu inference when device="auto", but I suggest we leave it off for now until a couple of fixes to improve inference speed on small datasets are in? This way the user has to explicitly turn it on, and has more of a chance of seeing if it's faster or not.

[gemini-code-assist]

Code Review

This pull request updates the docstrings for the device parameter in TabPFNClassifier and TabPFNRegressor to include information about multi-device inference. The changes are a good improvement, making the documentation clearer and more comprehensive.

I've suggested a minor wording change to generalize the documentation from 'GPUs' to 'devices', as the implementation seems to support parallel execution on various device types.

Additionally, I noticed that in src/tabpfn/regressor.py, the docstring for the device parameter is duplicated. The new, correct version has been added, but an old, conflicting version still exists further down in the docstring (lines 313-318). This should be removed to avoid confusion for users. Since this is outside the changed lines of this PR, I'm mentioning it here for your consideration.

[bejaeger]

Sounds good!