test: run the printer equivalence suite against Gecko in CI - #22155

`syntax-equivalence-firefox` runs the same corpus and the same assertions under
`EQUIVALENCE_BROWSER=firefox`, beside the Chrome job rather than replacing it —
a rewrite only one engine reads differently is invisible to a suite that asks
one, and five of the defects filed today are ones only Gecko sees. It is a job
of its own so the existing check keeps its name, installs the browser with
`yarn setup:firefox`, and uploads no coverage: the Chrome job already reports
this suite's lines under the `syntax-equivalence` flag.

`unsampledBy` counted a feature once per occurrence, so `(monochrome) and
(monochrome)` and the `(monochrome)` the printer folds it to read as two
conditions where the viewport bits agreed. `and`, `or` and `,` are idempotent,
so the parts are a set — its own documentation already said "named once".

The Chrome run was `syntax-equivalence` with nothing saying which browser it
asked, while the engine added beside it said `-firefox` — so the pair read as a
suite and an afterthought rather than as one suite over two engines. They are
one matrix now, `syntax-equivalence (chrome)` and `syntax-equivalence
(firefox)`, the shape `runtimes (bun)` and `runtimes (deno)` already use here.

That also removes the copy of the job the second engine had brought with it:
the corpus checkout, the three caches and both comparison installs were
duplicated. Firefox is installed on its own row, Chrome reports the coverage
(both cover the same lines, so a second uploader only leaves the report short),
and `fail-fast: false` keeps one engine's failure from cancelling the other.

A matrix row carries the script and the coverage flag beside the browser, and
the generated name prints every key — so the checks read `syntax-equivalence
(chrome, cover:syntax-equivalence, true)`. An explicit `name` keeps the engine
and drops the rest.

WebKit is the third engine and the one Safari ships, and puppeteer reaches
Chrome and Firefox only — so `EQUIVALENCE_BROWSER=webkit` drives it through
playwright, under a shim in `launchBrowser.js` presenting the nine calls the
suites make. Playwright takes one argument where puppeteer takes several, and
offers no `color-gamut`, so it declines media emulation outright and a
condition carries what no viewport varies as text, the way Gecko already does.

`forEngine` no longer assumes two engines: a reason opening `<engine> only:` is
kept where it names the engine being run and dropped everywhere else, so a
third filed set needs nothing further.

The row is `continue-on-error` until its filed set is written from what it
actually sees. Both other engines opened with hundreds of differences that were
nearly all the harness rather than the printer, and the same is expected here;
filing them before the first run would be inventing them.

The first WebKit run reported 24 differences over 5803 checks, and they are
five causes. Each is filed with what was measured, so the row blocks like
Chrome's and Firefox's rather than reporting past a CI flag.

Three are the harness meeting a third engine. WebKit reads a different
`document.compatMode` for the two forms of eleven files and the reverse for
two, which a direction that flips per file says is how the document is built
rather than what the printer wrote. It lays out in sixty-fourths of a pixel,
so `112px` reads back as `111.984375px` in five framework stylesheets. And it
serializes a computed number to six significant digits where Blink keeps more
— which is the printer's own cap, so the whole-element comparison now reads
both sides at that precision, Chromium excepted since the fixtures were
measured against it.

Two are the engine: no `fetchpriority` reflection, so `high` and `low` both
read `auto`; no CSS Painting API, so a `paint()` image is dropped; and
`shape-image-threshold` read as the number its percentage means.

`attribute-tables` and `minimize-values` are filed by two engines for
different causes now, so a reason may name more than one.