Make test_stem less flaky.

Since parametrizing the test allows it to run in parallel, this makes it
flaky, as one process can overwrite the test result image of another.

Our standard way for dealing with tests that use the same baseline image
is to pass duplicate filenames to `image_comparison`, because that is
serialized.

Author: QuLogic

lib/matplotlib/tests/test_axes.py

@@ -3342,10 +3342,15 @@ def test_hist_stacked_weighted():
     ax.hist((d1, d2), weights=(w1, w2), histtype="stepfilled", stacked=True)
 
 
-@pytest.mark.parametrize("use_line_collection", [True, False],
-                         ids=['w/ line collection', 'w/o line collection'])
-@image_comparison(['stem.png'], style='mpl20', remove_text=True)
-def test_stem(use_line_collection):
+@image_comparison(['stem.png', 'stem.png'], style='mpl20', remove_text=True)
+def test_stem():
+    # Note, we don't use @pytest.mark.parametrize, because in parallel this
+    # might cause one process result to overwrite another's.
+    for use_line_collection in [True, False]:
+        _test_stem(use_line_collection)
+
+
+def _test_stem(use_line_collection):
     x = np.linspace(0.1, 2 * np.pi, 100)
     args = (x, np.cos(x))
     # Label is a single space to force a legend to be drawn, but to avoid any