Temporal Logic Specification-Conditioned Decision Transformer for Offline Safe Reinforcement Learning