[Dummy] nan output by tensorrt

[Zha0q1]

This PR is used to reproduce the nan output issue when running inference of mxnet's bert onnx export on tensor rt in fp 16 mode #19747

Steps to export bert to onnx:

1. build mxnet and run pip install -e python/
2. install dependencies with pip onnx==1.6.0 onnx-simplifier onnxruntime==1.6.0 gluonnlp
3. cd export_bert
4. edit export_bert.py to toggle pretrained=True/False and specify batch and seq_length
5. python export_bert.py --layer=3 to export the model.
6. look into bert_model folder and use the *_sim.onnx onnx export file. This is the final output graph which has been optimized by onnx-simplifier

Steps to build trt engine from onnx file and run inference

1. use this image https://ngc.nvidia.com/catalog/containers/nvidia:tensorrt
2. pip install onnxruntime
3. copy build_trt.py and infer_trt.py and the onnx files into the container
4. edit build_trt.py to specify onnx input and trt output path and do python build_trt_py
5. edit infer_trt.py to specify onnx and trt path and batch and seq_length and do python infer_trt_py
6. step 5 will run both trt and onnxruntime and output the results by each

Provided trt files
I have also included two trt files that I generated by following the above steps so you can cut to step 5 above.
mx_bert_layer3_simp_1_16_pre.trt means BS 1 SeqLen 16 pretrained=True
mx_bert_layer3_simp_1_17_pre.trt means BS 1 SeqLen 17 pretrained=True

[mxnet-bot]

Hey @Zha0q1 , Thanks for submitting the PR
All tests are already queued to run once. If tests fail, you can trigger one or more tests again with the following commands:

• To trigger all jobs: @mxnet-bot run ci [all]
• To trigger specific jobs: @mxnet-bot run ci [job1, job2]

---

CI supported jobs: [centos-gpu, unix-cpu, sanity, clang, windows-cpu, centos-cpu, website, windows-gpu, miscellaneous, edge, unix-gpu]

---

Note:
Only following 3 categories can trigger CI :PR Author, MXNet Committer, Jenkins Admin.
All CI tests must pass before the PR can be merged.