ADD: support of fp16 python inference backend

[skprot]

Enviroment:

CUDA Version: 11.4
TensorRT version: 8.2.0.6-1+cuda11.4
ONNX version: 1.10.1
Python version: 3.8.10

Issue:

While working around with converted to onnx model with fp16 precision i have found the following issue:

Error Code 4: Internal Error (fp16 precision has been set for a layer or layer output, but fp16 is not configured in the builder)

This error occurs when building an engine in python when calling the command:
engine = backend.prepare(model, device='CUDA:0', verbose=True)

Then it was found that in the backend.py file there was no initialization of the fp16 flag in the config for building the engine.

Solution:

The solution is inspired by the still open pull request #345, but it's for the old version of TensorRT. The current solution works on the latest version of TensorRT, which at the time of the pull request is 8.2.0.6.

[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.

---

Sayan Protasov 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.}