…a:sort

Two related changes in the autosort comment-handling path:

1. Restore the pre-2.13.0 `has_comments` skip for config-driven autosort.

   Before the overhaul (emkguts#155/emkguts#158), Autosort.run/2 had this guard:

     should_skip || has_comments || !is_sortable -> {:cont, ...}

   i.e. `autosort: [:map | :defstruct | :schema]` did not reorder a
   collection that contained comments. emkguts#155 removed it and emkguts#158 replaced
   it with sorting the collection and trying to move comments with their
   keys. That works for per-key comments. For a standalone comment that
   heads a run of entries (the common real-world pattern in config,
   i18n, and large fixtures), sorting alphabetically disperses the run
   and the header ends up next to whatever key now follows it.

   The two patterns are structurally identical at the AST + comment
   level, so a heuristic can't distinguish "comment for one key with an
   uncommented neighbour" from "section header over many keys" without
   breaking the per-key case the overhaul intended to support.

   Adds `has_comments_inside?/2` back and includes it in the run/2 skip
   cond. `# quokka:sort` is handled separately (CommentDirectives calls
   Autosort.sort/2 directly) and is unaffected, so users who really want
   a commented collection sorted opt in per-value.

2. Sort the comments list returned by `sort_keyword_list_with_comments`
   by `.line`.

   `sort_groups_sequentially` prepends placed comments per iteration,
   so the returned list was in reverse line order. The else branch (no
   reorder needed) likewise returned `pair_comments ++ rest` unsorted.
   The formatter emits comments in list order, so on a `# quokka:sort`
   over a map with multiple per-key leading comments, the per-pair
   comments rendered reversed and clumped above one key, and on a
   second pass the directive comment was pulled inside the map. Sorting
   the returned list by line fixes both: leading comments stay with
   their keys after sort, and `# quokka:sort` stays above the map and
   is idempotent.

Test changes:
- Three existing tests that asserted config autosort reorders commented
  maps now assert the restored behaviour: the map is left untouched and
  mix format's end-of-line comment hoist still applies.
- New skip-regression tests for: map with a section comment over
  multiple keys, defstruct atom-list with a comment, defstruct keyword
  form with a comment, schema block with a comment.
- New `# quokka:sort` regression test asserts multiple per-key leading
  comments travel with their keys and the directive is preserved.

Docs/changelog updated to describe the restored rule and point users at
`# quokka:sort` as the per-value opt-in for commented collections.