Support remote inference on Triton + TensorRT or vLLM or TGI

[percyliang]

The preferred way to run models is to stand up an inference server (e.g., Triton + TensorRT or vLLM or TGI) locally and then hit it from HELM as an API. This way, HELM can benefit from all the crazy inference optimizations that are done. We need to demonstrate a proof of concept and write docs for this.

[yifanmai]

Opened draft PR #1975 for vLLM.

[yifanmai]

The TGI part is duplicated by #1866.
I don't know of any users asking for Triton currently, so I will deprioritize that.