Fix a bug in optimizer's mix_lr/max_lr when args.override_opt_param_scheduler==True

[lyuwen]

1. In the current setting, OptimizerParamScheduler uses the max_lr and min_lr from optimizer instead of its own attributes to compute the learning rate at a certain step.
2. When --override-opt_param_scheduler is used, user expects to override the learning rate schedule that is stored in the checkpoint. However, the optimizer still loads the max_lr, min_lr and the decoupled versions from the checkpoint.
3. Therefore, when --override-opt_param_scheduler is used, the learning rate in training is computed using the old max_lr and the new init_lr, a mixture of old and new setting. This can be confusing for user to figure out the issue.

Here I propose an fix that during the load_checkpoint, when --override-opt_param_scheduler is set to True, the function megatron.core.optimizer._update_min_and_max_lr_in_param_groups is invoked to update the optimizer loaded from the checkpoint with the new learning rate boudaries. Thus the training would be carried out using the updated learning rate setting in situations like CPT, etc.