Fix several recent issues from transformers, pandas, gpytorch

[BenjaminBossan]

1. Transformers v5: T5 now gives slightly different numerical results. I haven't investigated why, there were many changes in v5, it's probably not worth digging into. I just updated the tests to reflect the new results.
2. Transformers v5: The repr of tokenizer instances changed.
3. Transformers v5: hasattr(model, 'get_encoder') is no longer a reliable way to identify encoder-decoder models. Using model.config.is_encoder_decoder now. I checked that the attribute exists in transformers v4.37.2 (release 2 years ago) and it's there, so this should be a safe change.
4. Pandas v3: dtype-related error message changed. Note that I split a parametrized test into two tests as the error message now depends on the input dtype. Also, I had to take into account that the error message now depends on whether pyarrow is installed or not. Finally, as pandas 3 is not available for Python 3.10, I had to add a version check and change the error message conditionally.
5. GPyTorch v1.15: Some objects are no longer pickle-able. It's probably not fixable on the skorch side. For now, I pinned the version to <= 1.14 and opened an issue ([Bug] Can't pickle function CholLinearOperator.to_dense cornellius-gp/gpytorch#2717).

[thomasjpfan]

This error message is less obvious about what is wrong compared to "The following columns have dtypes that cannot be interpreted as"`. Is there a way to keep the original error message?

[BenjaminBossan]

That was a great suggestion, done.