Add generic deep research workflow: to pit against UDR and TTD-DR

[MrCsabaToth]

This would be pure LangGraph based, possibly maximum three rounds of research expansion and report development with judge logic. Then this can be vetted against UDR and TTD-DR, maybe side-by-side (#19).

[MrCsabaToth]

Some candidate to get inspiration from:

• First, since the NVIDIA Blueprint's VectorDB is Milvus, here's one from Zilliz: https://github.com/zilliztech/deep-searcher
• Collection: https://github.com/DavidZWZ/Awesome-Deep-Research
• ByteDance DeerFlow https://github.com/bytedance/deer-flow
• LiveResearchBench https://arxiv.org/html/2510.14240v1
• https://medium.com/@vi.ha.engr/modern-llm-frameworks-for-deep-research-open-source-vs-proprietary-solutions-642136c20078
• https://medium.com/@vi.ha.engr/modern-llm-frameworks-for-deep-research-open-source-vs-proprietary-solutions-642136c20078