Better `on_arch` sibling handling and tests

- Avoid a last-match-wins bug when multiple `on_arm`/`on_intel` pairs
  exist in different scopes. Grouping matches by parent lets us detect
  and autocorrect every eligible pair instead of only the final pair
  encountered.
- Add regression tests for multiple eligible `on_arch` pairs across
  sibling `on_os` blocks and for a later lone `on_arch` block after a
  correctable pair.

Co-authored-by: Issy Long <me@issylong.com>
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>