Serialize benchmark jobs to fix amdci3-1 rapid-dispatch corruption

[ChrisRackauckas-Claude]

Summary

Add job-level concurrency + max-parallel: 1 to the matrix so only one benchmark job runs at a time across all Benchmarks workflow runs. This works around a runner-pool bug where amdci3-1 corrupts its workspace under rapid dispatch.

Symptom

When 2+ benchmark jobs are dispatched to amdci3-1 within seconds, all but the last one die with infra errors. Reproduced three times now:

Date (UTC)	Cause	Jobs lost
2026-05-10	Original v7 merge (run 25374549824)	3 cancelled-mid-checkout
2026-05-16 07:58	First runner-free burst (PRs #1559/#1560/#1561)	6 lost, including StiffODE that had been running cleanly for 16h
2026-05-16 08:31	Single-folder PRs #1563–#1567 dispatched together	All 5 lost

Errors are always one of:

• ##[error]The operation was canceled. (during actions/checkout)
• ##[error]Missing file at path: ..._runner_file_commands/set_output_...
• ##[error]Can't find action.yml under .../_actions/actions/checkout/v6. Did you forget to run actions/checkout before running your local action?

amdci1-1 doesn't show this pattern. Single jobs on amdci3-1 also run cleanly (StiffODE, StiffSDE, AutomaticDifferentiation, NBodySimulator all completed normally when they had the runner to themselves).

Fix

benchmark:
  ...
  concurrency:
    group: benchmark-runner-pool
    cancel-in-progress: false
  strategy:
    matrix: ...
    fail-fast: false
    max-parallel: 1

Every matrix entry across every Benchmarks workflow run shares the constant group benchmark-runner-pool, so they queue one at a time. Other jobs (detect-changes, manual-matrix, publish) are unaffected.

Tradeoff

Even with both runners healthy and idle, only one benchmark job runs at a time. Halves throughput in the best case. Acceptable given the alternative is ~60% job-loss rate from rapid-dispatch corruption.

Can be relaxed once the underlying runner bug is fixed (e.g. parameterize the group by matrix.runner to allow one job per runner instead of one job globally).

Test plan

• Merge → next master push or PR triggers Benchmarks → matrix entries queue serially → all entries either succeed or fail on real bugs, none on infra
• Reopen retrigger PRs Retrigger StiffODE (single-folder, infra-failed in burst-dispatch) #1563–Retrigger ComplicatedPDE (single-folder, infra-failed in burst-dispatch) #1567 (already exist) — they should succeed one at a time

Note for review

Please ignore until reviewed by @ChrisRackauckas. This is an architectural choice, not a no-op fix.