Skill or Luck? Return Decomposition via Advantage Functions