Fix an issue with one_level_list schemas in parquet reader. (#10750)

Partially addresses: #10733

For a particular way of encoding list schemas (an old way that Spark seems to use sometimes), the parquet reader was accidentally propagating incorrect nesting information between columns.  Just a simple bug of not popping an extra value off a stack.

Note:  this is simply a fix so that the files read correctly, however the internal data in the file is actually of binary type and cudf converts these to string columns.  This PR does not add support for binary as a real type in cudf.

Authors:
  - https://github.com/nvdbaranec

Approvers:
  - Yunsong Wang (https://github.com/PointKernel)
  - MithunR (https://github.com/mythrocks)
  - GALI PREM SAGAR (https://github.com/galipremsagar)

URL: #10750

cpp/src/io/parquet/reader_impl.cu

@@ -679,6 +679,10 @@ class aggregate_reader_metadata {
           }
 
           std::copy(nesting.cbegin(), nesting.cend(), std::back_inserter(input_col.nesting));
+
+          // pop off the extra nesting element.
+          if (schema_elem.is_one_level_list()) { nesting.pop_back(); }
+
           path_is_valid = true;  // If we're able to reach leaf then path is valid
         }
 

python/cudf/cudf/tests/test_parquet.py

@@ -2386,6 +2386,32 @@ def test_parquet_reader_one_level_list(datadir):
     assert_eq(expect, got)
 
 
+# testing a specific bug-fix/edge case.
+# specifically:  int a parquet file containing a particular way of representing
+#                a list column in a schema, the cudf reader was confusing
+#                nesting information between a list column and a subsequent
+#                string column, ultimately causing a crash.
+def test_parquet_reader_one_level_list2(datadir):
+    # we are reading in a file containing binary types, but cudf returns
+    # those as strings. so we have to massage the pandas data to get
+    # them to compare correctly.
+    def postprocess(val):
+        if isinstance(val, bytes):
+            return val.decode()
+        elif isinstance(val, np.ndarray):
+            return np.array([v.decode() for v in val])
+        else:
+            return val
+
+    fname = datadir / "one_level_list2.parquet"
+
+    expect = pd.read_parquet(fname)
+    expect = expect.applymap(postprocess)
+    got = cudf.read_parquet(fname)
+
+    assert_eq(expect, got, check_dtype=False)
+
+
 @pytest.mark.parametrize("size_bytes", [4_000_000, 1_000_000, 600_000])
 @pytest.mark.parametrize("size_rows", [1_000_000, 100_000, 10_000])
 def test_parquet_writer_row_group_size(