[FEA] Zero copy output

[oyilmaz-nvidia]

Right now, once the output is generated by the python transform function, that output is copied to the Triton output buffer because Triton would like to control the lifecycle of the output buffer.

In C++, we can easily use this allocated output buffer by Triton but we'll check if this is possible with the python as well.

During one of the conversations with Triton team, this issue was raised and they said they can provide a callback function to control the third party output buffers. If we really need this kind of solution for python, we can check with the Triton team.