Fix test_sparse_norm_op failure. (#55405) (#55637)

test_sparse_norm_op的case在某些cudnn版本中，sparse和dense的batch norm计算结果存在较大diff。同样的结果可以在pytorch中复现，将数据dump之后用cuda api调用cudnn计算也可以复现。可以确认是cudnn版本导致的单测失败。因此暂时在单测中替换测试case，后续等nv有反馈后再对代码做进一步修改。

test/legacy_test/test_sparse_norm_op.py

@@ -100,19 +100,23 @@ class TestSparseBatchNorm(unittest.TestCase):
         else:
             bn = paddle.nn.BatchNorm3D(shape[-1], data_format=data_format)
         y = bn(x)
-        y.backward()
+        np.random.seed(5)
+        loss_data = np.random.uniform(-0.01, 0.01, y.shape).astype("float32")
+        loss = paddle.to_tensor(loss_data)
+        y.backward(loss)
 
         sp_x = paddle.to_tensor(data).to_sparse_coo(dim - 1)
         sp_x.stop_gradient = False
         sp_bn = paddle.sparse.nn.BatchNorm(shape[-1], data_format=data_format)
         sp_y = sp_bn(sp_x)
-        sp_y.backward()
+        sp_loss = loss.to_sparse_coo(dim - 1)
+        sp_y.backward(sp_loss)
 
         np.testing.assert_allclose(
-            y.numpy(), sp_y.to_dense().numpy(), rtol=1e-5
+            sp_y.to_dense().numpy(), y.numpy(), rtol=1e-5
         )
         np.testing.assert_allclose(
-            x.grad.numpy(), sp_x.grad.to_dense().numpy(), rtol=1e-5
+            sp_x.grad.to_dense().numpy(), x.grad.numpy(), rtol=1e-5
         )
 
     def test_nd(self):