Scheduler with respect to global_step (#844)

Also:
1. implemented __repr__ for the schedulers
2. Fix issue int/int. Some of the progress value is int64 torch tensor, which can cause problem for division for LinearScheduler.