feat(ci): parallelize qc into 3 jobs + composite setup action + caches

[w1ne]

Summary

Cuts CI wall-clock from ~3:30 to ~1:30 (≈45% reduction) by parallelizing the `qc` step into three independent jobs and adding three caches.

Spec: `docs/superpowers/specs/2026-05-04-ci-acceleration-design.md` (commit e460243).

What changes

• `package.json`: new `qc:lint` / `qc:build` / `qc:test` sub-scripts; existing `qc` becomes a meta-script that runs them sequentially. Local `npm run qc` workflow unchanged.
• `.github/actions/setup-cad/action.yml` (new composite): setup-node + node_modules cache + .tsbuildinfo/.vite cache + conditional npm ci. Shared across all jobs in both workflows.
• `.github/workflows/ci.yml`: three parallel QC jobs (lint / build-and-checks / test) + existing e2e job (now gated on all three). Standalone `Build` step removed (redundant with `qc:build`).
• `.github/workflows/deploy.yml`: same three QC jobs + new `web-build` job for the Vite production bundle (Pages artifact). e2e + deploy gates on the three QC jobs (deploy also gates on web-build).

Validation

• Spec: this PR's CI run IS the validation. Watch the Actions tab — three jobs should run in parallel, wall-clock should be max-of-three (~90s), not sum.
• Cache miss on first run is expected (fresh keys); cache hit on subsequent runs proves the cache restore path works.
• Local: `npm run qc` still runs all eight underlying commands sequentially in the original order.

Risks

5 risks documented in the spec; key ones:

• Cache miss rate too high — keyed on source content; partial-hit fallback via `restore-keys:` for the build cache.
• Build step removed — `vite build` regressions surface at release-time (deploy.yml) instead of PR-time. Acceptable trade.
• Composite action drift — kept tight (setup + caches + conditional install only); per-job extras live in jobs.

Test plan

• CI on this PR shows three parallel jobs in the Actions tab.
• Wall-clock < 2:00 on this PR.
• Subsequent push to this branch → cache hit, wall-clock < 1:30.
• All previously-passing tests still pass.