Embracing Unknown Step by Step: Towards Reliable Sparse Training in Real World