ci: split ci.yml into reusable build + test workflows

[robtaylor]

What & why

The organically-grown ci.yml had reached 2056 lines with build and test
logic interleaved per backend. This splits it into a build → test pipeline
behind a thin orchestrator, so every target configuration is compiled once on a
cheap standard runner and the resulting binary is handed to the test jobs on the
appropriate runners. Full design in docs/plans/ci-build-test-split.md.

ci.yml (orchestrator) ── changes ── build.yml ── test.yml
                                  └─ docs (Pages, ungated)

• build.yml (reusable, workflow_call): a 4-way matrix producing one
  artifact per config — jacquard-{cpu,cuda,hip,metal}, all carrying the
  synth on-ramp. CUDA is all-major SASS so one build feeds both the T4 and
  the self-hosted Blackwell runner.
• test.yml (reusable): the full test batch run against those prebuilt
  artifacts. CPU/compile-in-place jobs on ubuntu-latest; CUDA/HIP on the
  GitHub-hosted T4; Metal + MCU-SoC + JTAG on macos-latest-xlarge.
• ci.yml is now a 129-line orchestrator. It passes code (docs-only
  gate) and build_result into test.yml so a failed build surfaces as a
  failed required check (via each artifact job's explicit build-gate step)
  rather than a silently-passing skip.
• Divergent toolchain setup and the artifact download+chmod are factored into
  composite actions under .github/actions/ (setup-{cuda,hip,metal},
  init-submodules, download-jacquard), shared between build and test.

Scope: faithful move

Each backend keeps its current test content (Metal its rich suite, CUDA/HIP
the shared gpu_test_suite.sh). Unifying the batch across backends is a tracked
follow-up, not part of this PR — it would run new tests on CUDA/HIP for the
first time and needs GPU-runner triage.

Decisions baked in

• synth is uniform on every artifact (published + test binaries).
• Runner strategy: functional gating on GitHub-hosted runners (parallel);
  self-hosted nvidia1 reserved for perf. cuda-blackwell is non-gating —
  it skips (not fails) when the build didn't succeed, so an offline box never
  blocks a merge.
• CVC MCU-SoC path is an advisory guide, not a gate: mcu-soc-comparison
  gains continue-on-error so a CVC↔Metal mismatch is surfaced in the step
  summary without blocking merges.

Validation

Green on the first run — every job passed, including the artifact reuse
(one CUDA build → T4 + Blackwell), the new cosim-cpu artifact-consume path,
and backend equivalence:
https://github.com/gpu-eda/Jacquard/actions/runs/28785097079

⚠️ Required before merge: update branch-protection check names

The test jobs now run inside the reusable test.yml, so their check names gain
a Test / prefix. The 7 currently-required checks must be renamed:

Old (remove)	New (add)
Lint	Test / Lint
Unit Tests	Test / Unit Tests
Metal Tests (macOS)	Test / Metal Tests (macOS)
CUDA Tests	Test / CUDA Tests
HIP Tests (NVIDIA backend)	Test / HIP Tests (NVIDIA backend)
Benchmarks	Test / Benchmarks
MCU SoC Metal Simulation	Test / MCU SoC Metal Simulation

Because this PR reports the new names, it can't satisfy the old required
checks — merge it via admin override, then immediately update the required set
(other in-flight PRs will pick up the new names on rebase).

Follow-ups (separate PRs, build on this)

1. Unify the test batch across backends (gpu_test_suite.sh).
2. release.yml reuse of build.yml → publish Linux CUDA binaries; move
   release-metal off self-hosted macos-runner-1.
3. Single fat CUDA+HIP binary (jacquard-only; ulib already supports it).
4. Real AMD/ROCm runner on nvidia1.