On the Robustness of Open-World Test-Time Training: Self-Training with Dynamic Prototype Expansion.