Trt refactor remove mixin

[andompesta]

This is the main PR used to refactor TRT inference code to:

• remove onnx tracing dependencies
• support multiple trt-configuration for the same model
• trt-build done using polygrafy subprocess for easier understanding

[andompesta]

merge 8 description

This PR is in charge of removing onnx-exporter class as it is not needed.
Onnx model comes from HF repo.

[andompesta]

merge 9 description

This PR is in charge of adding a trt-config class.
This class is responsible to build the trt engine given an onnx path and various trt-flags.

As the same model might need different trt-configurations depending on which precision is used, a registry is used to collect all the model configuration.
Based on the provided key the get_config method will return the appropriate model configuration to use.

@timudk here are the main changes
Each configuration is a dataclass containing:

• the needed trt flags
• a from_model factory method to feed all needed parameters to the config class
• an get_input_profile method that return the max and mix input supported by the build engine
• build process is still based on polygraphy, but instead of using python API a subprocess CLI approach is used for clarity

Engine classes are changed to:

• use trt-config class instead of mixin and exporter classes

[andompesta]

merge 10 description

This PR implements the changes on trt-manager to use trt-config classes instead of exporters and mixin.

for each porvided model a trt-config is provided by _get_trt_configs method.
for each trt-config a trt-engine is build. trt-build is fully based on trt-config classes. There is no more exporter/model_config split
after trt-engines are build, a trt-runtime is initialized
finally engine classes are instantiated with a valid cuda-stream

@timudk this PR greatly simplify this class by reducing code lines by half.
I'm keeping trt-config and engine classes separate for clarity. They could be merged in a single, really big classes, which I believe it would not provide much benefit

[andompesta]

merge 11 description

This PR change the CLI script to support the new TRT interface.

1. A try ... except approach is used to solve tensorrt import: people that do not install trt dependencies would experience an error otherwise.
2. Instead of using

trt: bool = False,
trt_transformer_precision: str = "bf16",

a different set of inputs are used:

• trt_onnx_dir is used to specify the folder containing onnx models
• trt_engine_dir specify the location where trt engines will be loaded from or build
• trt_precision specify the precision of the pipeline. For now the supported precision are: bf16, fp8 and fp4
  All of these 3 args are needed to run trt engines. I found this approach to be easier rather than using a mix of input args and env-variables. @timudk please advise on which approach would be preferred on your side.

The following are additional inputs args that can be provided:

• trt_batch_size the batch size used to optimize the engine, by default 1
• trt_static_batch weather the engine is build with static batch size, by default yes
• trt_static_shape weather the engine is build to support static image shape, by default yes

Note that:

• Based on the provided trt_precision a set of input values are generated for the trt-manager class.
• if TRT is not available, but input arguments are provided: an error is raised
• all trt related input arguments start with trt- prefix