Beyond Non-Expert Demonstrations: Outcome-Driven Action Constraint for Offline Reinforcement Learning