Segmentation fault if run `tests/mxnet/test_nn.py` on cpu with python 3.7.0

## 🐛 Bug

it works well with python 3.6.9. seems related to `dgl.add_self_loop()` in below code: https://github.com/dmlc/dgl/blob/39987bc59970289b774319f9e7ccd1ad57a1cb98/tests/test_utils/graph_cases.py#L92-L97

Related PR: https://github.com/dmlc/dgl/pull/4406
## To Reproduce

Steps to reproduce the behavior:
`docker run -it dgllib/dgl-ci-cpu:v220816 /bin/bash`
`conda activate mxnet-ci`
`clone/build/install DGL`
`DGLBACKEND=mxnet PYTHONPATH=tests:$PYTHONPATH python -m pytest tests/mxnet/test_nn.py::test_graph_conv2`

## Expected behavior

## Environment

- DGL Version (e.g., 1.0): master, 0.9
- Backend Library & Version (e.g., PyTorch 0.4.1, MXNet/Gluon 1.3): mxnet 1.6.0
- OS (e.g., Linux): Linux
- How you installed DGL (`conda`, `pip`, source):
- Build command you used (if compiling from source):
- Python version: 3.7.0
- CUDA/cuDNN version (if applicable):
- GPU models and configuration (e.g. V100):
- Any other relevant information:

## Additional context