Add a /benchmark PR command for the adapter accuracy benchmark

[AlonMalach]

What

A maintainer's /benchmark comment on a pull request benchmarks the PR's head commit with the adapter accuracy benchmark, the way /gpu-test runs the GPU tests:

/benchmark                          the first model (granite-4.1-3b)
/benchmark --model granite-4.2-3b   another model
/benchmark --no-cache               re-run a commit that already has its row

The run publishes the commit's row to https://generative-computing.github.io/granite-switch/benchmarks/, sets a commit status, and posts the results on the PR.

Why only these two files

GitHub runs comment-triggered workflows only from the default branch. The benchmark itself (the harness, its workflow adapter-benchmark.yaml and the results page) lives on feature/adapter-benchmark and does not touch main. This PR adds only the launcher:

• .github/workflows/benchmark-command.yaml reacts to /benchmark comments on a GitHub-hosted runner.
• .github/scripts/benchmark_command.sh checks the commenter's role (Maintain or Admin, with the existing check_role.sh), reads the flags, and dispatches adapter-benchmark.yaml on feature/adapter-benchmark with the PR's head SHA.

Both follow gpu-test-command.yaml and gpu_test_command.sh: the same role check, comment text passed only through env and quoted arguments, the first line only, and unknown flags declined rather than ignored.

Security

• The launcher's role check is fast-fail UX only. The dispatched workflow re-checks the role on the GPU runner with /opt/gsw/check_role.sh, like the GPU tests.
• Like /gpu-test, this runs the PR's code on the cluster, so a maintainer should read a PR before commenting.
• The workflow that runs on the GPU runner comes from feature/adapter-benchmark, so that branch should be protected like main.

Testing

• The handler, run locally with a stub gh: valid commands dispatch to the benchmark branch with the right inputs; a mistyped flag (--nocache), /benchmarks and a non-maintainer are declined, with nothing dispatched.
• The dispatched side already works: gh workflow run adapter-benchmark.yaml --ref feature/adapter-benchmark benchmarked b7e5421 on both models and published the rows.
• The comment trigger itself can only run after merge, since GitHub reads it from main.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!