improves trtllm example

[rcano-baseten]

:rocket: What

Addresses some shortcomings in the trtllm exmaple

• corrects forward incompatibility caused by ModelInput(**model_input) invocation
• more cleanly carves out type conversion for oai
• provides temporary fix for trt llm "stream=False" shortcomings, that results in generating nothing (cc @htrivedi99 @aspctu )

  :computer: How

:microscope: Testing

Deployed this on a Llama-3-8b-trt-llm model and ran it through its paces (stream=True, False), additional params to ensure forward compatibility, printed model input from truss to verify functionality

Some Follow Ups:

• update llama-3-trt-llm example code
• Anything else?

[rcano-baseten]

closing in favor of https://github.com/basetenlabs/truss/pull/1010