fix: make fingerprint feature deterministic

[LennartPurucker]

Using Python's built-in hash is not consistent between runs due to salting.

Thus, the fingerprint feature using the built-in hash is not deterministic, making the input data non-deterministic and making the predictions non-deterministic between runs. Hashlib's hash is consistent between runs and thus does not face this problem.

This broke during the refactoring. The commit below is essentially reverting a commit that happened during the refactor.

FYI: @eddiebergman

[LeoGrin]

Thanks Lennart, LGTM! I would maybe just add a test that TabPFN is deterministic when setting the random state? I thought that scikit-learn checks would already check that but they were passing before. (I checked quickly and the only failing test test_sklearn_compatible_estimator[TabPFNRegressor()-check_methods_sample_order_invariance] still fails with this PR)

[LennartPurucker]

We cannot test for this, as it requires separate Python runs (not just separate calls).

And test_sklearn_compatible_estimator[TabPFNRegressor()-check_methods_sample_order_invariance] fails due to precision and not randomness.

[LeoGrin]

Ah yeah ok the test does look hard to do! LGTM then!

[eddiebergman]

Ah damn, apologies. @LeoGrin here's the reference Lennart's refering to:
https://docs.python.org/3/reference/datamodel.html#object.hash

Scroll to the bottom of that function