Bootstrapping Adaptive Human-Machine Interfaces with Offline Reinforcement Learning