DiSA-IQL: Offline Reinforcement Learning for Robust Soft Robot Control under Distribution Shifts