Lyapunov-Guided Self-Alignment: Test-Time Adaptation for Offline Safe Reinforcement Learning