Marginally reduce memory footprint of joins (#14197)

If we drop the gather maps as soon as we are done with them, we have a little more headroom for joins that are close to hitting the device memory limit.

Authors:
  - Lawrence Mitchell (https://github.com/wence-)

Approvers:
  - GALI PREM SAGAR (https://github.com/galipremsagar)
  - Bradley Dice (https://github.com/bdice)
  - Matthew Roeschke (https://github.com/mroeschke)

URL: #14197

python/cudf/cudf/core/join/join.py

@@ -203,6 +203,7 @@ def perform_merge(self) -> cudf.DataFrame:
             if left_rows is not None
             else cudf.DataFrame._from_data({})
         )
+        del left_rows
         right_result = (
             self.rhs._gather(
                 GatherMap.from_column_unchecked(
@@ -213,7 +214,7 @@ def perform_merge(self) -> cudf.DataFrame:
             if right_rows is not None
             else cudf.DataFrame._from_data({})
         )
-
+        del right_rows
         result = cudf.DataFrame._from_data(
             *self._merge_results(left_result, right_result)
         )