Merge: Use Custom instead of OHE encoded categorical parameter for Non-SHAP tests (emdgroup#732)

There seem to be machine-dependent non-deterministic failures for the
LIME method which is part of our SHAP insight (as non-SHAP extension
handled mostly via shap).

The failures do not indicate any particular problem on our side but some
internal process causing NaN's when calculating model regression
metrics. The test only seems to fail when there is the `Categorical_1`
parameter involved, which is a 3 category, OHE encoded parameter. This
was replaced with the custom parameter as it avoids degenerate values
which hopefully contribute to avoiding the NaN generations in LIME.

I've [tried this fix 3 times in the `Regular`
action](https://github.com/emdgroup/baybe/actions/runs/21355269316) (had
to ignore one pip-audit vulnerability to do so in a temporary commit
which has been removed again). All runs had no issue, which I take as
indication the fix worked. If the issue appears again I would probably
add another parameter to the test searchspaces to avoid generation of
degenerate points.

Author: Scienfitz

.github/workflows/regular.yml

@@ -140,8 +140,7 @@ jobs:
         py-version: [ {semantic: '3.10', tox: 'py310'},
                       {semantic: '3.11', tox: 'py311'},
                       {semantic: '3.12', tox: 'py312'},
-                      {semantic: '3.13', tox: 'py313'},
-                      {semantic: '3.14', tox: 'py314'} ]
+                      {semantic: '3.13', tox: 'py313'} ]
     name: Full Tests ${{ matrix.py-version.semantic }}
     runs-on: ubuntu-latest
     steps:

tests/insights/test_shap.py

@@ -55,8 +55,8 @@ def batch_size():
         ["Conti_finite1", "Conti_finite2"],
         ["Num_disc_1", "Fraction_1"],
         ["Num_disc_1", "Conti_finite1"],
-        ["Num_disc_1", "Categorical_1"],
-        ["Conti_finite1", "Categorical_1"],
+        ["Num_disc_1", "Custom_1"],
+        ["Conti_finite1", "Custom_1"],
     ],
     ids=[
         "params_conti",