Constraint-Adaptive Policy Switching for Offline Safe Reinforcement Learning