Track agent workflow gaps from block histogram review

Follow-up from #9222.

During review of #9222, reviewers had to point out several CCCL/CUB style and convention issues that should have been caught before review. This issue tracks a post-mortem/follow-up PR to capture durable guidance in the appropriate place, rather than adding broad agent/benchmarking documentation to #9222 itself.

Initial scope:
- Review the #9222 feedback and classify misses as documented in `AGENTS.md`, inferable from nearby code, or tribal knowledge.
- Move/adapt the proposed `AGENTS.md` and `docs/cub/benchmarking.rst` guidance into a dedicated docs/conventions PR if still appropriate.
- Keep the histogram algorithm PR focused on the CUB implementation, tests, benchmarks, and algorithm docs.

The follow-up PR should be created from an issue-named branch and can close this issue once the guidance is reviewed.