feat: expose LightRAG concurrency and extraction knobs in settings

[ademicho123]

Description

Adds three settings LightRAG didn't expose: max_concurrent_files (file-processing concurrency), llm_model_max_async (LLM call concurrency), and entity_extract_max_gleaning (extraction depth). They're wired through the same lightrag.json settings file and /rag-pipelines/lightrag/config endpoint that already handle top_k/response_type, so the pattern should look familiar. Under the hood, max_concurrent_files goes to RAGAnythingConfig, and the other two ride into LightRAG's own constructor via RAG-Anything's lightrag_kwargs passthrough.

Scoped down from the full ask in #640: fallback models and single-provider lock-down live in the shared LLM client, not this pipeline, so I left those for a separate issue/PR rather than mixing them in here.

Related Issues

• Related to [Feature Request]: Specify parameters of LightRAG system #640

Module(s) Affected

• api
• config
• services
• tests

Checklist

• I have read and followed the contribution guidelines.
• My code follows the project's coding standards.
• I have run pre-commit run --all-files and fixed any issues.
• I have added relevant tests for my changes.
• I have updated the documentation (if necessary).
• My changes do not introduce any new security vulnerabilities.

Additional Notes

Added 7 tests: defaults/clamp coverage for the three new settings (tests/services/config/test_graphrag_lightrag_settings.py), unit tests for the two new settings-reader functions covering both happy path and read-error fallback, and two tests confirming build_rag() actually threads the values into RAGAnythingConfig/RAGAnything (tests/services/rag/test_lightrag_pipeline.py). Also had to patch the fake classes in the existing #594 regression test, since it broke once build_rag() started passing the new kwargs through — full suite is green aside from one pre-existing, unrelated Windows path-separator issue in that same #594 test.

Open to feedback on the default values (max_concurrent_files=1, llm_model_max_async=4, entity_extract_max_gleaning=1) if you'd rather they match something else.