ci: persist ccache across ephemeral runner pods via actions/cache

[MrLi000001]

Reminder

• I have read AGENTS.md and docs/developers_guide/agent_governance.md.
• I have linked an issue or explained why this PR does not need one.
• I have added adequate unit tests and/or case tests, or explained why not.
• I have listed the exact verification commands run and their results.
• I have described user-visible behavior changes, including INPUT parameter changes.
• I have explained core-module impact for ESolver, HSolver, ElecState, Hamilt, Operator, Psi, or other source/ changes.
• I have requested any needed governance exception below.

Linked Issue

This PR does not need an issue: it is a CI-infrastructure-only change. The motivation is
backed by 1683 workflow runs (2026-04 to 2026-08): ARC runner pods rebuild from an empty
ccache on every job (build median 2021s / ~34min) while VM runners hit their warm local
cache (median 155s) — a 13x gap that is growing as the pod share rises (16% in June to
39% in August). The pod cache is lost because ephemeral pods are recreated per job and
/tmp/ccache lives in the pod's temporary storage.

Unit Tests and/or Case Tests for my changes

• Commands run:
  • Local YAML parse of both modified workflow files:
    python -c "import yaml; yaml.safe_load(open('.github/workflows/test.yml')); yaml.safe_load(open('.github/workflows/cuda.yml'))"
  • Local zstd pack/unpack roundtrip on the ccache directory (tar + zstd -T0).
• Result summary: YAML validation passed; pack/unpack roundtrip preserves ccache -s stats.
• Checks not run, with reason: no local full build — this is a workflow-only change and the
  GitHub Actions run itself is the verification; expected CI behavior (documented below)
  will be confirmed by the run logs after merge.

What's changed?

• Persist the ccache directory of test.yml / cuda.yml through the GitHub Actions Cache
  Service so ephemeral ARC runner pods no longer start every job with a cold cache.
• Restore the cache only on ARC pods (if: startsWith(runner.name, 'gpu-runner-')); VM
  runners keep their warm local cache untouched and only upload it as the cloud seed.
• Pack ccache as a zstd tarball before upload (~4.3G -> ~600M) to stay within the 10GB free
  cache quota; unpack after restore with ownership fix.
• Cache key: ccache-<toolchain>-${{ github.sha }} with
  restore-keys: ccache-<toolchain>- — exact key for re-runs, prefix fallback for new
  commits/PRs. runner.name is intentionally absent from the key because pod names are
  random per job (would always miss) and VM/pod caches should be shared.
• Set CCACHE_BASEDIR: ${{ github.workspace }} so cached objects remain portable across
  checkout paths.
• Print RUNNER_NAME / runner type in the logs to allow timing statistics grouped by
  runner type (pod vs vm).
• Expected CI behavior: run Merge recent abacus changes to deepmodeling #1 after merge = seed (restore miss, full build, upload);
  subsequent runs = restore hit, pod build step from ~2021s down to the minute scale.

Governance Notes

• INPUT/docs changes: none.
• Core module impact: none — workflow-only change.
• Exceptions requested: none.

---