Feature request: Custom instructions for LLM-based transcription models

[igor-gorohovsky-sembly]

We now have models which can act like both LLM and Transcriber, so we can specify post-processing instructions and audio input in one request (Voxtral example). It would be quicker than making two requests.