refactor: remove redundant per-repository work and make output testable

[bircni]

What

Three kinds of cleanup, all behaviour-preserving.

Redundant work removed

• Status::new walked the index and working directory twice for every
  dirty repository — once to decide clean-vs-dirty, then again through
  get_changed_count to reach a number it had already seen. It now counts
  in the pass it already makes.
• is_git_worktree was strictly narrower than is_git_directory (a
  worktree's .git is a file, and exists() is true for files), so
  a.is_git_directory() || a.is_git_worktree() was only ever a. Removed,
  along with the duplicated stat calls per walked directory.
• The remote was resolved up to three times per repository.
• The unchanging scan root was canonicalized once per repository.
• A String was allocated for every walked directory and discarded on the
  two most common paths.

Structure

• find_repositories collected into Arc<RwLock<Vec<_>>> and then deep
  cloned both vectors out from under a held read guard. It now collects with
  rayon and partitions — dropping the locks, the Arc, the clones, and the
  parking_lot dependency.
• The scanning closure is split out as scan_entry; the error it previously
  discarded unlogged is now logged like its sibling branch.
• RepoInfo::new took three consecutive positional bools, with nothing at
  the call site preventing a swap. It now takes &Args.

Output is now testable

run accepted a writer it only used for completions — every printer function
wrote to println!. The writer is now threaded through, so run's signature
is honest and the tests assert on what is actually rendered. The printer tests
were smoke calls with comments standing in for assertions (// Assert that the table is printed correctly); they now check real output. printer.rs goes
from largely unverified to fully covered.

Compatibility

Output is unchanged. A binary built from the parent commit and one built from
this branch produce byte-identical output across the flag matrix, on a fixture
covering clean, dirty, detached, local-only, unpushed, stashed, worktree,
nested, unreadable and subdir repositories.

Two duplicate tests were removed, and the wall-clock assertions in the smoke
tests were relaxed from 2s to a deliberately generous bound — they exist to
catch a pathological regression, not to fail on a loaded CI runner.

Deliberately not included

• Merging get_ahead_behind_and_local_status with get_branch_push_status.
  They overlap, but use different notions of upstream (branch.upstream() vs
  a literal refs/remotes/{remote}/{branch} lookup) and can legitimately
  disagree — unifying them changes behaviour and deserves its own PR.
• Moving find_repositories out of cli.rs into a scan module.
• Removing the derived has_unpushed field — it is part of the JSON output.