Use structured output for conversation generation

[brumar]

This would prevent the LLM to disrespect the <Person> markups in its output. This would also make https://github.com/souzatharsis/podcastfy/issues/5 a bit more comfortable (user prompts would not break or undermine the markup instructions).

Link for gemini: https://ai.google.dev/gemini-api/docs/structured-output?lang=python