* Do not terminate a hack-prefixed property before a comment

A declaration is read as a custom property when the *first token* of the
declaration starts with `--`, but the stringifier decided it from `prop`.
The `*`/`_` hack prefix is moved out of `prop` into `raws.before` after that
decision, so for `*--x:red` the parser builds a normal declaration -- whose
value stops at the first comment -- while the stringifier saw a custom
property and terminated it:

  postcss.parse('a{*--x:red/*c*/}').toString()
  // => 'a{*--x:red;/*c*/}', semicolon invented

The same applies to any declaration with something other than spaces in
`before`. Check `before` alongside `prop` so the two stay in sync: only a
declaration that will re-parse as a custom property can swallow a following
comment, and only that one needs the semicolon.

* Shorten the comment

---------

Co-authored-by: Jayesh <[email protected]>

split() only pushed an item at a separator when it had collected some
text, so an empty value was dropped unless it happened to be the last
one. list.comma(',,') returned [''] instead of ['', '', ''].

Whitespace hid the inconsistency: 'a, ,b' collects ' ', which is not
empty, so it survives and is then trimmed to ''. That makes the result
depend on the spacing rather than the number of commas:

    list.comma('a,,b')  // ['a', 'b']
    list.comma('a, ,b') // ['a', '', 'b']

Rule#selectors reads through list.comma, so 'a,,b' reported two
selectors and assigning them back rewrote the rule as 'a,b'.

The `last` flag already marks the separator as significant, which is
what comma() passes and space() does not, so it is the right condition
here too. Runs of whitespace still collapse for space().

Signed-off-by: maximilliangrand <[email protected]>
Co-authored-by: maximilliangrand <[email protected]>

list.split() tested current for emptiness before trimming it, so a chunk made only of whitespace that is not one of the separators survived the check and was pushed as an empty string. Trimming first makes the emptiness check see the value that actually gets pushed.

clearAnnotation() has two implementations. The AST one removes only a
comment whose text starts with `# sourceMappingURL=`, and
PreviousMap#loadAnnotation finds annotations by the same marker. The
string one, which is the only path NoWorkResult can take, searched for
`/*#` and removed every comment that began with it.

So postcss().process(css) silently deleted `/*#region */` and
`/*#endregion */` folding markers, and any other comment starting with
`#`, while the same CSS through a processor with one no-op plugin kept
them. It happens with `map` unset too, because NoWorkResult calls
clearAnnotation() unconditionally.

Searching for the whole `/*# sourceMappingURL=` marker lines the string
path up with the other two and keeps the plain string scan.

Co-authored-by: Dylan Pulver <[email protected]>