feat: answer generation step to async

[Youssouf-Boudouia]

The proposed patch lack feature such as request per minute management and doesn't give meaning-full result on my benchmark, since the quantity of test already present is not enough.

The way I choose to implement the patch is to change the code here and there in order to not introduce breaking change. We can do better for maintainability, especially for controlling 'async pollution'

I think it is better to use the 'Routing Strategy' API from little because it provide asynchronous management and request per minute configuration.

In overall, I may not fully understand the reason under the current implementation, whoever I'm convinced there is different approach that ease the batch processing implementation.

As I understand, the goal is to provide a tool that ease the testing comparative against different LLM. In this perspective I think creating a declarative API how allow configuration in the pipeline is a must.

My recommendation is to define what kind of data the system need and treat them as configuration, and in a second time how the package should be used. This will allow users to define the necessary data and configuration parameters in a clear and intuitive manner, streamlining the setup process and enhancing overall user experience.

[GillesMoyse]

Just tested it in a specific branch, it is working well! We nonetheless need to add a limit to the number of calls per minute before merging.