docs: add integrations roadmap and progress tracker

[Davidobot]

Summary

Adds docs/integrations.md, a living tracker for LodeDB's integrations into the open-source RAG / agent ecosystem — a single source of truth for external-integration work.

Contents

• Status table — LangChain (shipped), LlamaIndex (in review, feat(integrations): add LlamaIndex VectorStore adapter #7), mem0 (next), PrivateGPT, Haystack, txtai, AnythingLLM, Open WebUI, Flowise, kotaemon, LocalGPT, Dify (evaluating), Khoj (blocked — AGPL).
• Sequencing — Now (LlamaIndex) → Next (mem0, then PrivateGPT) → Later (framework breadth + visibility apps).
• Planned developments — per-target notes, with mem0 detailed (provider interface, config slot, delta-persistence fit) and the rationale for each.
• Design constraints — text-path embedding (and a vector-in API as the future option), exact-match-only metadata filters, the Python/Rust ↔ JS/TS language boundary (local bridge for JS/TS apps), and GPU/CUDA scope.
• Shared benchmark plan — BEIR (SciFact / FiQA / NFCorpus) + synthetic 100k/1M, the metric set, and fairness rules (fixed embedding/chunking, brute-force recall oracle, always include batched + concurrent modes since the GPU crossover is ≈ batch 50).
• Licensing notes — permissive-first; Khoj AGPL is blocked; Open WebUI / Dify source-available caveats.

Targets and sequencing draw on integration research across both the framework and app surfaces. Star counts and licenses are marked approximate and dated, to be re-verified before external use.

Docs-only change. References #7 (the LlamaIndex adapter).