fix(lint): treat an HTML comment as a text boundary outside inline elements

[VXNCXNX]

An HTML comment between two words causes a false spelling alert:

$ cat test.md
on <!-- TODO --> a
sonata

$ vale test.md
test.md:2:2  error  Did you really mean 'ona'?  Vale.Spelling

Removing the comment yields no alerts.

lintHTMLTokens sets closedInline in the EndTagToken branch and clears it in StartTagToken, but a CommentToken leaves it untouched. Its consumer is clean(), whose whitespace decision is (closedInline && spaced) || (!closedInline && inline && !starter). After a comment both flags are stale-false, so neither arm fires, the following text loses its leading space, and on + a fuse into ona.

Setting closedInline on a comment only fixes the outer case: a comment right after an inline opening tag is not a closed boundary, and asserting one there disables the #1052 padding. That loses real alerts and invents words — A sona<em><!--c-->ta</em> fuses to sonata, so the genuine sona alert disappears.

So the flag is set only when no inline element is open. Inside one, the opening tag remains the boundary and its padding is preserved.

Fixes #882

Verification

Built before/after binaries and compared:

input	before	after
on <!-- TODO --> a / sonata	2:2 'ona'	no alert
A sona<em><!--c-->ta</em> here.	1:3 'sona'	1:3 'sona'
A sona<b><!-- c --></b>ta here.	unchanged	unchanged
in <code><!--c-->x</code> for	unchanged	unchanged
word<!--c-->s	stays words	stays words
A <!-- c --> sona	no alert (fused to Asona)	1:14 'sona'

The last row is a second bug the same change fixes: text after a comment at the start of a fragment was being fused to the preceding word.

go build ./... and go test ./... pass. The regression case is in testdata/e2e/comments.yaml alongside the existing #1001 comment case, and it fails without the fix.

Note lintHTMLTokens is shared across html/md/adoc/rst/org/dita/xml, so this affects all of them; I checked the .html path behaves the same.

[jdkato]

Thanks!