[BUG] lr scheduler not changed when resume and args is changed

[triwahyuu]

Describe the bug When resume training and the scheduler args is changed, the scheduler behavior is not changed.

To Reproduce

1. train model with this lr_scheduler config:

   lr_scheduler: {
   method: StepLR,
   args: {step_size: 200}
   }

2. resume training with this config:

   lr_scheduler: {
   method: StepLR,
   args: {step_size: 350}
   }

3. the learning rate still stepped when hit epoch 200.

Additional context this happened because the lr scheduler load the state dict from the resumed model, should be changed.