Performance fixes for yolox-tensorrt demo

[awarebayes]

Hello! I did some performance tweaks for the yolox Tensorrt demo.

They are the following:

1. Check whether the box has high enough confidence, and only then calculate its dimensions
2. Make grid_and_strides static, in order for it not to recalculate on every call to decode_outputs. This takes around 50-60% of runtime when I perform inference on video, and can be calculated just once and then reused

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

Mikhail Scherbina seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[awarebayes]

I'd been using it for video inference and have found these changes usefu

[awarebayes]

It says I have agreed to the CLA for Megvii-BaseDetection/YOLOX on the website, but the pipeline does not update

[FateScript]

It says I have agreed to the CLA for Megvii-BaseDetection/YOLOX on the website, but the pipeline does not update

Did you sign the CLA with your github account ? What if you try another web browser ? @awarebayes