[mkldnn-v1.0] Skip flaky test for unidirectional rnn_relu

Skip `test_rnnrelu_sym`, and add some issue tracking message

tests/python/unittest/test_operator.py

@@ -38,6 +38,7 @@
 def check_rnn_consistency(cell1, cell2, T, N, I, H, grad_req, rtol=1e-2, atol=1e-4):
     if default_context().device_type == 'cpu':
     # NOTE(zixuanweeei): Currently, we don't add `add` requests support on fused mkl-dnn rnn operator.
+    # We tracked this issue by https://github.com/apache/incubator-mxnet/issues/16578
         if isinstance(grad_req, dict) and 'add' in grad_req.values():
             print("Skip the test when requiring `add` operation against gradients on CPU context.")
             return
@@ -257,6 +258,8 @@ def test_rnntanh_bidirectional():
 @with_seed()
 @assert_raises_cudnn_not_satisfied(min_version='5.1.10')
 def test_rnnrelu_sym():
+    if default_context().device_type == 'gpu':
+        print("Skip test `rnn_relu_sym` on gpu. This is tracked by https://github.com/apache/incubator-mxnet/issues/16548")
     Ts = [1, 5]
     Ns = [1, 32]
     Is = [32, 128, 512]