Use parquet rather than checkpoint to break lineage

The reason for this is that checkpointing can cause the schema
to be lost if there are zero rows.  Whereas parquet preserves the schema

tests/helpers.py

@@ -74,7 +74,7 @@ def db_api_args(self):
         return {
             "spark_session": self.spark,
             "num_partitions_on_repartition": 1,
-            "break_lineage_method": "checkpoint",
+            "break_lineage_method": "parquet",
         }
 
     def convert_frame(self, df):