Resa: Transparent Reasoning Models via SAEs