* exclude lr scheduler's state from accumulators_holder

* fix when there is no learning rate scheduler

* make a copy of the loaded state dict to avoid modifying it