ci(benchmark): run the engine benchmark on GitHub Actions and post the report to Discussions #2125

Is your feature request related to a problem? Please describe.
The engine benchmark (examples/benchmark/run.sh) only runs by hand on a local machine. So the docs' Engine Support speed table goes stale between releases. #2109 had to re-measure it once 3.15 shipped.

Describe the solution you'd like
A benchmark.yml workflow:

• Triggers: a weekly schedule, at an off-hour slot clear of the Saturday crons and of Monday's CI Timing and Python Versions, plus workflow_dispatch with inputs for the Python versions and the rounds.
• Steps: run examples/benchmark/run.sh on ubuntu-latest, which is native x86_64 with Docker, then upload out/ as an artifact.
• Report: a discussion job, modelled on ci-timing.yml's, the only job with discussions: write. It posts the ratio table (table.rst), with its absolute-time footnote, through util/discussion_report.py to a benchmark thread in CI reports, naming the runner image.
• Docs: stay hand-updated from the report; the workflow never pushes.

Describe alternatives you've considered
Running per push or PR: the runner pool is already saturated. Auto-committing the docs tables: that bypasses review, and the absolute ms come from shared, unknown hardware. The within-pass ratios cancel most runner noise; the absolute numbers do not.

Additional context
Runtime of a full matrix on a hosted runner is unmeasured. The first dispatch measures it, and that sets timeout-minutes and the cadence.