fix: enforce max_workers in LLMMetadataExtractor.run_async

[etairl]

Summary

• LLMMetadataExtractor.run_async acquires its asyncio.Semaphore once around the outer gather(...) instead of inside each task, so max_workers has no effect and every prompt in a batch fires its LLM call simultaneously.
• The docstring on __init__ advertises max_workers as "the maximum number of requests that should be allowed to run concurrently when using the run_async method", so the current behavior silently breaks that contract and can blow up LLM-provider rate limits / connection pools on large batches.
• Fix is to move the async with sem: into a per-task wrapper coroutine so the limit is actually enforced. Added a regression test that verifies peak in-flight calls stay <= max_workers.

Before

sem = Semaphore(max(1, self.max_workers))
async with sem:
    results = await gather(*[self._run_async(prompt) for prompt in all_prompts])

After

sem = Semaphore(max(1, self.max_workers))

async def _bounded_run(prompt: ChatMessage | None) -> dict[str, Any]:
    async with sem:
        return await self._run_async(prompt)

results = await gather(*[_bounded_run(prompt) for prompt in all_prompts])

Test plan

• hatch run test:unit -k test_llm_metadata_extractor passes (includes new test_run_async_respects_max_workers).
• CI green.

[vercel]

@etairl is attempting to deploy a commit to the deepset Team on Vercel.

A member of the Team first needs to authorize it.