AI provider spans are suppressed globally when LangChain integration is active

[nicohrubec]

When both LangChain and an AI provider integration (e.g. OpenAI) are active, we globally disable provider instrumentation to avoid duplicate spans. This means direct provider SDK calls that aren't related to LangChain also lose their spans. We should suppress provider spans only within the context of a LangChain call, not globally.

[linear-code]

JS-1884

[nicohrubec]

We have a draft PR open to solve this issue, but will put off merging for a bit to evaluate whether this really accounts for all our requirements: #19688