Safe reinforcement learning with online filtering for fatigue-predictive human-robot task planning and allocation in production