test(css): compare the css-parsing-tests trees rather than only running them - #22243

`should install onto terser` asserted that under Deno `phases` is empty,
because Deno could not import terser's own sources when the test was
written. It sometimes can, and then the test fails — it went red once on
#22243, a pull request touching only CSS fixtures, and cost a re-run of a
21-minute job to establish it was not that change's doing.

Which sources a runtime reaches is the one thing `load` is written to
tolerate: it catches any failure and hands back terser's published entry
point. So the runtime is the wrong thing to assert. What `load` owes
everywhere is that it installed phases this build declares and nothing
else, and Node, which always reaches the sources, still owes the whole
set — both of which this now states.

The fallback path stays covered by behaviour rather than by a hardcoded
outcome: every `should mangle exactly as terser does` case compares
webpack's `minify` against terser's own, whichever way `load` resolved.

* fix(css): read an rgb()'s separator grammar before folding it

`_minifyColorFunction` cuts on whitespace, `,` and `/` in one pass without
recording which separator stood where, so every spelling collapsed to the
same argument list: `rgb(1, 2 3)`, `rgb(1 2, 3)`, `rgb(1 2 3, .5)`,
`rgb(1, 2, 3 / .5)`, `rgb(1 2 3 .5)` and `rgb(1 / 2 / 3)` all folded to a
color. Each is a declaration an engine drops, so folding it changed what
rendered.

This is the defect #22242 fixed for `hsl()`, `hwb()` and the Lab family,
left out of that change because `rgb()` reads its arguments through a
scanner of its own. The same two grammars decide it — the modern `A B C`
with an optional `/ D`, and the legacy `A, B, C` with an optional `, D`,
both of which `rgb()` has — so the check is the pair of regexps that change
already added, applied after the scan rather than before it, where only a
plausible color reaches it.

The pending changeset for #22242 covers this, so it is widened rather than
duplicated.

* chore(changeset): keep the color entry within the length limit

The union of the two color fixes ran to 148 characters, where
`.changeset/**` asks for one imperative sentence of 80 or fewer. Both
facts still fit: the unitless channels that now fold, and the separator
grammar that decides whether anything folds at all.

* test(javascript): assert what load() owes, not what Deno can import

`should install onto terser` asserted that under Deno `phases` is empty,
because Deno could not import terser's own sources when the test was
written. It sometimes can, and then the test fails — it went red once on
#22243, a pull request touching only CSS fixtures, and cost a re-run of a
21-minute job to establish it was not that change's doing.

Which sources a runtime reaches is the one thing `load` is written to
tolerate: it catches any failure and hands back terser's published entry
point. So the runtime is the wrong thing to assert. What `load` owes
everywhere is that it installed phases this build declares and nothing
else, and Node, which always reaches the sources, still owes the whole
set — both of which this now states.

The fallback path stays covered by behaviour rather than by a hardcoded
outcome: every `should mangle exactly as terser does` case compares
webpack's `minify` against terser's own, whichever way `load` resolved.

…22252)

`should install onto terser` asserted that under Deno `phases` is empty,
because Deno could not import terser's own sources when the test was
written. It sometimes can, and then the test fails — it went red once on
#22243, a pull request touching only CSS fixtures, and cost a re-run of a
21-minute job to establish it was not that change's doing.

Which sources a runtime reaches is the one thing `load` is written to
tolerate: it catches any failure and hands back terser's published entry
point. So the runtime is the wrong thing to assert. What `load` owes
everywhere is that it installed phases this build declares and nothing
else, and Node, which always reaches the sources, still owes the whole
set — both of which this now states.

The fallback path stays covered by behaviour rather than by a hardcoded
outcome: every `should mangle exactly as terser does` case compares
webpack's `minify` against terser's own, whichever way `load` resolved.