fix: normalize Arrow-backed arrays to numpy in Ray test comparison

Ray 2.56.0 changed to return Arrow-backed DataFrames from Ray Data batches,
so df[col].values returns ArrowExtensionArray instead of numpy ndarray.
numpy.all() on bool[pyarrow] triggers a pandas/pyarrow incompatibility
(axis kwarg not accepted). Use Series.to_numpy() to always get numpy arrays.

Also revert the incorrect pandas<2.3 pin (wrong root cause diagnosis).

Author: w4nderlust

pyproject.toml

@@ -30,7 +30,7 @@ classifiers = [
 ]
 dependencies = [
     "numpy>=1.24",
-    "pandas>=2.0,<2.3",
+    "pandas>=2.0",
     "scipy>=1.10",
     "tabulate>=0.9",
     "scikit-learn>=1.3",

tests/integration_tests/test_ray.py

@@ -265,8 +265,11 @@ def _to_hashable(v):
         df1 = df1.iloc[key1.argsort(kind="stable").values].reset_index(drop=True)
         df2 = df2.iloc[key2.argsort(kind="stable").values].reset_index(drop=True)
     for column in df1.columns:
-        vals1 = df1[column].values
-        vals2 = df2[column].values
+        # Use to_numpy() rather than .values to ensure numpy arrays regardless of
+        # backing store. Ray 2.56+ returns Arrow-backed DataFrames from Ray Data
+        # batches, so .values may return ArrowExtensionArray instead of numpy.
+        vals1 = df1[column].to_numpy()
+        vals2 = df2[column].to_numpy()
 
         if any(feature_name in column for feature_name in [CATEGORY]):
             is_equal = np.all(vals1 == vals2)