Rethinking Open-World Semi-Supervised Learning: Distribution Mismatch and Inductive Inference