[TrtLLM] Support JIT compilation and dynamic batch for TrtLLM python backend

[sindhuvahinis]

Description

T5 handler support is in another PR. Separated my changes into multiple PR for easy reviewing #1680

This is for TensorRT-LLM T5 python backend

• Does JIT compilation of T5 models
• Automatically sets dynamic batch if not provided.
• Automatically sets engine=MPI and mpi_mode=True

NOTE: User has to disable rolling_batch in order for us to set dynamic batch and engine=mpi with these changes. We could do disable rolling batch ourselves, but did not do that way because, if TRTLLM release C++ backend support for T5 in between our releases, user could upgrade their TRTLLM in requirements.txt and enable rolling batch and test them. But open to suggestions here.

Testing:
Tested with flan-t5-xl in my EC2 machine.

My serving.properties

option.model_id=/opt/ml/model/t5/hf-model
option.rolling_batch=disable
option.entryPoint=djl_python.tensorrt_llm

[lanking520]

why we cannot use the old param?

[sindhuvahinis]

Because that param checks whether rolling_batch == trtllm. Might as well, remove the param.

case "rolling_batch":
      isTrtLlmBackend = "trtllm".equals(value);
       break;

https://github.com/deepjavalibrary/djl-serving/pull/1678/files#diff-244860c398daaa2ff4bc328b98c867ccce93643b1f65c4f1024bf6bdd5a7beecR124