Inconsistent behavior between CPU and GPU on ReLU operator when input is NaN

[maybeLee]

Describe the bug I was converting a keras to ONNX through one library (tf2onnx)[https://github.com/onnx/tensorflow-onnx], however, I find out that: when giving a NaN input, onnxruntime will correctly output NaN in CPU mode while output normal value in GPU mode. After some debugging, I find out this inconsistency happens when I set the activation of dense to be relu. Specifically, the simplest graph that can trigger this bug is as follows:

To Reproduce

• Describe steps/code to reproduce the behavior.
• Attach the ONNX model to the issue (where applicable) to expedite investigation. The ONNX model can be accessed here You can directly run this code to generate the ONNX model and reproduce this issue: https://colab.research.google.com/drive/1UGjTV1LcV_YZye1v1mAIQet-qwhDP1s9?usp=sharing

Expected behavior Same as executing on CPU mode, onnxruntime should also output NaN when executing on GPU mode

Screenshots

[cloudhan]

I think that is caused by fmaxf in cuda

If both arguments are NaN, returns NaN.
If one argument is NaN, returns the numeric argument.