per page processing for better speed

### Requested feature

Currently doclings process document pipeline E2E iterates all pages. Some threading can be configured.

When inferencing models via i.e. vllm and having a machine with 8 gpus and initially 8 documents - things are fine. But some documents are longer than others and this means eventually 7 gpus will be idle - and the 1 remainig (data skew) cannot fully utilize the others.

I would love to find a way forward how to handle data skew in docling in a resource efficient way.

Obviously this is a bit more complicated: Some VLLM hosted models behave differently ad scale to more gpus. I think the question is also how to harmonize this scaling behaviour in a generic way.

No file, test, or concrete entry point is named. Start by tracing the document pipeline E2E and its threading configuration, then reproduce the uneven page-length workload with vLLM across eight GPUs. Done requires an agreed generic approach for per-page processing and data-skew handling, with evidence that resource utilization improves.

Written by the indexing model from the issue text.