Move creation of prompts to a function

[DzmitryPihulski]

Create a separate function build_all_requests() to use in all inference functions. It will return all the prompts, possibly with applied chat template.