Update LLM and AsyncLLM to expose more functionality

[gyulaz-htec]

Changed LLM class to provide a callback which can be used to insert additional requests to a running LLM evaluation

• This can be used to better balance the load between multiple TP1 instances for offline inference.
  Updated AsyncLLMEngine with the following:
• Added option to only send back first token(s) and the final result
• Replaced execute_model_async with execute_model. According to rpd trace results this has a lesser CPU (few ms) overhead.