Event processor migration: Vercel AI event processor

[Lms24]

core/src/tracing/vercel-ai/index.ts

Another tough one: The vercel AI event processor does a lot of post processing:

• setting/removing attributes on a single ended span (processEndedVercelAiSpan, applyToolDescriptionsAndTokens) should be easy to port to client.processSpan
• We accumulate token usage data on parent spans (accumulateTokensForParent, applyAccumulatedTokens). This can become more dangerous because to guarantee continued correctness, we need a guarantee that the child spans end before their parent so that we can accumulate at end time. This might need careful investigation if its the case. Otherwise we loose accuracy.

[linear-code]

JS-2214

[Lms24]

A note on this: After talking with @alexander-alderman-webb: Instead of continuing to accumulate the token usage stuff, maybe we can switch already to only record the attributes on the child span level and avoid setting anything on parent spans. This should reduce the complexity for streamed spans.