Nvidia Triton model packaging feature

[karbyshevds]

As Data Scientist I want to be able to pack (containerize) trained mode to Nvidia Triton or other server and deploy it using ODAHU deployment.
Deployed model should be able to accept tensor, now it can only accept JSON that should be de-serialized before passing data to model.
Triton provides ability to pass raw data to models and such features as micro-batching and GPU sharing so it also may help to save costs.