ci: run the invariant checks once instead of per matrix leg

[Tinitto]

The server-and-driver matrix has five legs — mock, aer, quantify, qblox, sim — and every one of them ran the same go vet, the same ruff, and the same CLI suite. Four redundant copies of each. lint-go depends on build-dashboard, so each leg also built the dashboard twice: once as an explicit step and once as that prerequisite. Ten dashboard builds a run, of which five did nothing.

Those three checks move to a lint-and-test-shared job that runs once. The matrix keeps only what actually varies by extra: the dependency set, the extra's own Python tests, and its driver end-to-end run.

The dashboard build stays on the legs that run a driver end to end, because build_server is a bare go build and main.go embeds the dist — but not on sim, which runs no driver.

Both publish gates gain the new job, so a lint failure still blocks a release.

Also wires in make test-dashboard. The target has existed since the graph work but CI invokes targets individually rather than through make test, so the dashboard's unit tests have never run there. It goes ahead of the Cypress step: a broken helper should fail in seconds, not as a puzzling assertion against a live server four minutes later.

Not addressed here: the sim leg pays for the whole thirty-three-node DAG walk twice, once per scheduler, at ~90s each. Measured, that is 78% of test_calibration_loop.py, and it is evenly spread across the routines — the slowest is ramsey at 3.7s — so there is nothing to trim without weakening the tier-3 fits. Splitting the two schedulers into their own legs would halve that wall clock, and is worth measuring once this lands.

[ch-ahindura]

LGTM