Add LightRAG integration

[gitbuda]

• Make it work with a fast local LLM
• Profile the execution -> per page/small_doc it's slow ~10-12s/page, local gpt-oss-20b on legacy RTX 3060 takes ~85s, multiple LLM calls take most of the time

---

• Make a robust setup to run local LLM
• Add option to enable/disable calculating embeddings
• Parallelize the LLM reqs -> there is the built-in batching https://github.com/HKUDS/LightRAG/tree/main?tab=readme-ov-file#insert (use the max_parallel_insert arg)

[antejavor]

Have you planned to have this released on PyPI?

[gitbuda]

Merging this.
The future is listed under the #88 (I think we should, once it's polished, publish the LightRAG integration under the PyPi).