bench(extract-worker): perform load testing

[ClemDoum]

In order to avoid discovering our system at large scale the first time we'll need it, perform load testing at moderately large scale to assess:

• end-to-end processing time (with temporal overhead, machine provisioning and so on): "how many hours does it take to extract markdown content from P pages"
• does it scale properly
• can we find a fast/lightweight/cheap config vs. precise/full VLM config
• how much does it cost ?
• is worker startup fast enough for small volumes of documents ?
• cache model download on workers: https://docling-project.github.io/docling/usage/advanced_options/#model-prefetching-and-offline-usage

[ClemDoum]

Potential prority for sprint 50: P0, in progress in last sprint