There were two issues in skorch with sklearn v1.8 that are now fixed:

__sklearn_is_fitted__ should return a bool. Instead, we were just
raising or returning None. Not sure if this was already wrong when
added (#1119) or if the API changed later.

Moreover, sklearn was now raising an error when using the SkorchDoctor
class because of missing __sklearn_tags__. This class now inherits from
BaseEstimator, resolving this issue.

For good measure, I also tested locally with sklearn 1.7 to ensure that
these changes are backwards compatible.

Unrelated changes:

Remove an unnecessary import and trailing whitespace. Added
__sklearn_is_fitted__ to SkorchDoctor too.

Quick patch release for sklearn v1.8 compatibility.

Transformers fixes:

- Some numerical differences from T5
- Tokenizer repr has changed

Pandas 3: error message changed

GPyTorch: fix version to avoid pickle error

E.g. should now work with XPU.

- drop 2.6.0
- bump 2.9.0 to 2.9.1
- add 2.10.0
- update docs

The argument use_auth_token was deprecated and removed, use token
instead.

- add 2.11.0
- drop 2.7.1
- update triton install for CI
- GPyTorch pickling error is resolved, update tests

As the frequency of new PRs on skorch isn't very high anymore, it can
happen that the unit tests don't run at all for longer stretches of
time. This means that we may catch issues only quite late, and only if a
new PR lands do these unrelated errors surface (#1137 being a recent
example)

With this PR, the test unit tests on a weekly basis. The
schedule is Sunday morning, the same as for the HF integrationt tests.

Enable sklearn metadata routing in skorch.

See: https://scikit-learn.org/stable/metadata_routing.html