fix: avoid gluing a slash onto a comment in dialects where // starts a line comment

[Jeremy-xuan]

fix: avoid gluing a slash onto a comment in dialects where // starts a line comment

What's wrong

Snowflake is the only dialect where // starts a line comment. With
denseOperators on, a / operator directly in front of a comment gets glued to it,
forming // — so the rest of the line becomes comment text:

SELECT a / /* c */ b FROM t;

SELECT
  a//* c */ b
FROM
  t;

b is now inside the comment. The statement quietly means something else, and it
does not look wrong enough to catch by eye. A /* sql-formatter-disable */ marker
is swallowed the same way, so the user's "do not touch this" instruction stops
working:

SELECT a / /* sql-formatter-disable */ b FROM t;   --  ... a//* sql-formatter-disable */ b ...

Root cause

Layout.add() already had this exact guard for --:

  // Don't glue a layout item starting with "-" directly onto one ending with
  // "-": that forms "--", which re-parses as a line comment and
  // swallows the rest of the line (e.g. densing "a - -b" into "a--b").
  if (item.startsWith('-') && this.lastItemEndsWith('-')) {
    this.items.push(WS.SPACE);
  }

The reasoning is right, but it is hard-coded to --. Every other line comment
marker a dialect uses has the same hazard, and // is one of them.

The fix

The guard now asks the dialect for its own line comment markers instead of assuming
--. The markers already exist on the tokenizer options, so they are carried through
the same path as identifierDashes:

• ProcessedDialectFormatOptions gains lineCommentTypes
• dialect.ts fills it from tokenizerOptions.lineCommentTypes ?? ['--']
• Layout takes it as a constructor argument and checks it in wouldFormLineComment()

  private wouldFormLineComment(item: string): boolean {
    const lastItem = last(this.items);
    if (typeof lastItem !== 'string') {
      return false;
    }
    return this.lineCommentTypes.some(
      (marker) =>
        marker.length > 1 && lastItem.endsWith(marker[0]) && item.startsWith(marker.slice(1))
    );
  }

Behavior for -- is unchanged — the existing test does not glue a "-" in front of another "-" in dense mode still passes — and dialects that do not use // are not
affected. lastItemEndsWith() was left with no callers, so it is gone.

Tests

Three tests in test/snowflake.test.ts: a block comment after /, a
sql-formatter-disable comment after /, and a plain a / b that must keep being
densed (so the fix does not disable dense operators wholesale). The first two fail on
master.

pnpm test   →  Test Suites: 27 passed, 27 total
               Tests:       5849 passed, 1 skipped, 5850 total
pnpm run lint / pretty:check / ts:check / build   →  all clean

Regression check

master versus the fix over 12,605 combinations (6 dialects × 3 option sets × ~700 SQL
literals harvested from the test suite), with word-multiset preservation against the
input as the criterion: 7,688 byte-identical, 0 regressions. Only Snowflake
changes, and only where a / meets a comment.