Refactor the LLM benchmark script

[nv-hwoo]

Refactor the benchmark script for more modular codebase as well as to give users more flexibility on specifying input tensors or sampling parameters through either input data JSON file or command line when running benchmarks.

Before the change, if users wanted to sweep through generated prompts, they had to manually specify parameters (e.g. max tokens) through command line without providing input data JSON file. This was to avoid conflicts in the code. However, this limited users to use input data JSON file only when they were using custom prompt.

Now, users can use both input data JSON file and command line to specify parameters regardless of whether or not they want to sweep through prompts or use custom prompts. For instance, users can use

• input JSON only to set sampling parameters and generate prompt
• both command line & input JSON to set sampling parameters and generate prompt
• command line only to set sampling parameters and generate prompt
• input JSON to fix prompt (no prompt generation)

*When using both command line and input JSON file to specify parameters, if any parameter overlaps, the value specified in the command line will overwrite the one in the input JSON file.

[debermudez]

couple questions around your format calls. otherwise looks good.