[Skills] Add kernel benchmark sanity references

[WoosukKwon]

Summary

• add rough Blackwell sanity-check references for memory-bound, GEMM, and
  attention kernel benchmarks
• clarify that the figures are diagnostic reference points rather than gold
  standards, guaranteed peaks, or hard limits
• note the FlashAttention-4 source for the B200 BF16 attention result

Duplicate-work check

Searched open PRs for kernel-microbenchmark, PFLOP/s, and
FlashAttention-4 skill; no matching PR was found. This change is not linked
to an issue.

Tests

• .venv/bin/python /home/woosuk/.codex/skills/.system/skill-creator/scripts/quick_validate.py .agents/skills/kernel-microbenchmark — passed (Skill is valid!)
• git diff --check — passed
• .venv/bin/pre-commit run --files .agents/skills/kernel-microbenchmark/SKILL.md — passed (applicable hooks, including markdownlint and typos)

Model evaluation is not applicable because this changes contributor skill
guidance only and does not affect model output, accuracy, or serving behavior.

AI assistance

OpenAI Codex assisted with editing, duplicate-work checks, validation, the
commit, and drafting this PR description. The human submitter reviewed every
changed line and is responsible for the change.

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}