[feature request] Add `finish_reason` to LLM output message attributes**

[uu9]

Is your feature request related to a problem? Please describe.

finish_reason is not captured in the OpenInference semantic conventions or any instrumentation library (e.g. openinference-instrumentation-openai, openinference-instrumentation-langchain). This makes it impossible to observe why a model stopped generating — e.g. distinguishing stop, length, tool_calls, or content_filter — which is critical for debugging truncation issues, tool call flows, and safety filtering.

Describe the solution you'd like

1. Add finish_reason to the semantic conventions spec as a new attribute on output messages:

   llm.output_messages.{i}.message.finish_reason  String  e.g. "stop", "length", "tool_calls"
   

2. Collect it in openinference-instrumentation-openai from choice.finish_reason when processing ChatCompletion and Completion responses.

3. Collect it in openinference-instrumentation-langchain from the generation_info field in LangChain run outputs.

Describe alternatives you've considered

Monkey-patching _ResponseAttributesExtractor._get_attributes_from_chat_completion to inject finish_reason manually — works but is fragile and requires users to maintain the patch across library upgrades.

Additional context

OpenAI's choice.finish_reason is a standard field in all Chat Completion responses. The OTel GenAI semantic conventions already track this via gen_ai.response.finish_reasons. Aligning OpenInference with this would improve interoperability and observability coverage.

[mikeldking]

@uu9 this makes sense - any particular instrumentation you are looking for it in?