fix(test): return Series (not ndarray) from _parse_expected_datetime_column

pandas 2.x's Series.dt.to_pydatetime() returns an ndarray, tripping
mypy's return-type check against the pd.Series annotation and
(without dtype='object') letting pandas coerce the datetime.datetime
objects back to pd.Timestamp when Series wraps them — which
reintroduces the nanosecond overflow this helper exists to avoid.

Wrap with dtype='object' and preserve the source index.

Signed-off-by: mokashang <shangmengjiajiajia@gmail.com>

Author: mokashang

sqlmesh/core/test/definition.py

@@ -1032,7 +1032,11 @@ def _parse_expected_datetime_column(series: pd.Series, target_type: type) -> pd.
         return parsed.dt.date
     if target_type is datetime.time:
         return parsed.dt.time
-    return parsed.dt.to_pydatetime()
+    # `Series.dt.to_pydatetime()` returns an `ndarray` in pandas 2.x. Wrap it in a
+    # Series with ``dtype=object`` so pandas does not coerce the values back to
+    # ``pd.Timestamp`` (which would reintroduce the nanosecond overflow this
+    # function exists to avoid).
+    return pd.Series(parsed.dt.to_pydatetime(), index=parsed.index, dtype="object")
 
 
 def _normalize_df_value(value: t.Any) -> t.Any: