openai: handle developer role for openai messages

[xrmx]

Wait for resolution of: open-telemetry/semantic-conventions#1877

[codefromthecrypt]

fwiw I think that the role should be passed through, as there are other roles besides what is mentioned. I can look them up, but basically I think we don't want to cause maintenance or drop messages due to enum fitting.

[codefromthecrypt]

here's an example of the many potential roles, and this only includes ones openai puts in their swagger https://github.com/openai/openai-openapi/blob/master/openapi.yaml#L13372-L13376

I think the balance is that if we use roles as a hint, we don't have to use duck typing to figure out what sort of event structure we should make. OTOH there are many roles and this instrumentation is used in other platforms that may not precisely track openai's changes, and due to deployments even if they did it may be a long time before converging.

So, I think the choices might be:

1. pass through the role, instead of using it to lookup which fields to parse. Rather look at all fields and decide the structure of the event based on the structure of the response (e.g. tools even if there are data required in the response)
2. track openai's roles, but add a catch block which does 1. so that we don't drop events
3. just track openai's roles

I like 1 or 2 as I don't think we win by being a sort of schema enforcement system. OTOH there's grey area here. Hope these thoughts help!

[xrmx]

In the meantime that something will change in the semantic convention follow what java is doing (elastic/elastic-otel-java#539) and implement sending the system event with an explicit developer role in the body as the semantic conventions allow.