ci: cut wasted minutes from measurement_id and WASM jobs

[a10y]

Summary

Three targeted fixes from profiling recent CI runs (job/step timings + logs from PR runs of Linters and Tests):

• measurement_id golden vectors: 6.2 min → ~20 s. The setup-uv action defaults to uv sync --all-extras --dev, which builds the vortex-data PyO3 extension via maturin on an uncached ubuntu-latest runner — ~6 min of the job (verified in job logs: Run uv sync --all-extras --dev → Building vortex-data). The test itself takes ~1 s and uses uv run --no-project, so it needs no workspace packages. Pass sync: false, matching the existing python-cuda-test job.
• WASM integration smoke test: ~5.2 min → expected ~1–2 min. It was the only Rust-compiling job left on plain ubuntu-latest with no compiler cache; ~4.5 min is a cold cargo build. Moved to the prebuilt runs-on image with S3 sccache, following the same pattern as the other build jobs (fork fallback to ubuntu-latest preserved; rustup target add wasm32-wasip1 mirrors what the wasm32 build job does).
• Windows sccache observability. Rust tests (windows-x64) is the ci.yml PR critical path (consistently 6.8–7.7 min across the last 8 PR runs, ~5 min of it compile). The sccache server demonstrably starts, but logs give no way to see hit rates. Added an sccache --show-stats step (if: always(), main repo only) so cache effectiveness on Windows is visible per run.

Not included (infra-side observations from the same investigation)

• Codspeed PR runs take 9–14 min but shards compute for only 0.8–4 min — they queue 5.5–9.8 min for amd64-medium runners while concurrent ci.yml jobs get runners in <1 min. One PR push requests ~26 runs-on VMs at once; this looks like a concurrent-instance/pool cap in .github-private (this repo's runs-on.yml only defines images). Raising capacity there, or consolidating the 8 Codspeed shards to 4, would cut the largest remaining chunk of PR wall clock.
• The Windows pool uses m8i-flex.2xlarge (8 vCPU); a size bump would directly shrink the PR critical path since compile dominates and linking 53 test binaries is uncacheable.

Checks

• yamllint --strict -c .yamllint.yaml .github/workflows/ci.yml — passes.
• Workflow-only change; no Rust/Python code touched, so no cargo/pytest checks were run per repo guidance. The WASM job change is best validated by this PR's own CI run.

🤖 Generated with Claude Code