[R] Datasets — pin standard eval datasets

[AmitoVrito]

Goal

Use community-standard datasets for credibility.

Sub-tasks

• BEIR for retrieval
• HotpotQA / NQ for multi-hop
• AgentBench / WebArena for agents
• MT-Bench / Arena-Hard for LLMs
• License audit

Definition of done

• Dataset registry with versions + checksums