Adhere to semantic convention for the "finish reason" attribute in LLM observations.

Comment From: tzolov

LGTM, thanks @ThomasVitale