ci: use T4 + xlarge runners — re-enable CUDA/HIP, de-serialize Metal

[robtaylor]

Summary

Puts the two new gpu-eda team-plan runners to work alongside the free self-hosted macos-runner-1:

• tesla4-runner — 4 vCPU + 1 NVIDIA T4 (GitHub-hosted)
• macos-runner-xlarge — M2 Pro, 5-core CPU / 8-core GPU, 14 GB RAM/storage (GitHub-hosted)

What changes

CUDA + HIP CI back online (on the T4, every push). Both jobs were if: ${{ false }} and pinned to the offline nvidia-runner-1 since 2026-05-01 — CUDA has had zero CI coverage since. They now run on tesla4-runner:

• CUDA Tests: native on the T4.
• HIP Tests (NVIDIA backend): the job builds with hipcc + HIP_PLATFORM=nvidia, so the T4 validates the HIP code path too. Native AMD/HIP still needs an AMD/ROCm runner (future).

This directly restores backend coverage and unblocks #104 (CUDA/HIP sim-timing parity).

Metal de-serialized (xlarge, gated to main/label). The three Metal jobs used to serialize on the single self-hosted runner (the reason we stacked #91→#110→#111). The two light jobs (Metal Tests, JTAG Minimal Cosim) now have a conditional runs-on:

runs-on: ${{ (github.ref == 'refs/heads/main' || contains(github.event.pull_request.labels.*.name, 'ci:metal-xl')) && 'macos-runner-xlarge' || 'macos-runner-1' }}

• Routine PR pushes: stay on free macos-runner-1 (full coverage, no cost).
• main / a ci:metal-xl-labelled PR: offload to the billed macos-runner-xlarge, running in parallel with the disk-heavy MCU SoC Metal Simulation job — which stays pinned to macos-runner-1 because xlarge has only 14 GB storage (big designs won't fit).

Concurrency group added (cancel-in-progress per ref) so rapid pushes don't pile up on the self-hosted / billed runners.

Cost posture

T4 runs on every push (GPU coverage was the big gap); the billed macOS xlarge is gated to main/label so routine PRs cost nothing extra on macOS. Added the ci:metal-xl label for opting a PR into xlarge.

⚠️ Confirm before merge

• Runner label strings. I used the names exactly as given: tesla4-runner and macos-runner-xlarge. These must match the runner names registered in org settings. Note GitHub's standard Apple-Silicon larger-runner labels are macos-latest-xlarge / macos-15-xlarge — if macos-runner-xlarge isn't a custom larger-runner you named, the gated jobs won't schedule. Easy to adjust if so.
• First CUDA/HIP run may be red. These haven't built in CI since May and main has moved ~45 commits; the first run is diagnostic, not a regression from this PR. They aren't required status checks, so they won't block merges.

Stacking

Stacked on #111 (base feat/bidir-tristate-readback). Cascade: #91 → #110 → #111 → #112. Each auto-retargets toward main as the one below merges.

Registered both labels in .github/actionlint.yaml; actionlint clean (remaining findings are pre-existing shellcheck).