add ema to BaseTrainer init

[IliasChair]

This is a follow-up to PR #909

The original changes were made directly on the main branch of my fork. To continue working on my fork, it was necessary to move these changes to a dedicated feature branch and resubmitting them through this PR. The code changes remain the same.

Apologies for the mess!

Running user-generated models with the OCPCalculator currently raises an error:
'OCPTrainer' object has no attribute 'ema' at line 621 in the load_checkpoint function of base_trainer.py.

This issue occurs because self.ema is not declared when load_checkpoint is called. The ema attribute is currently declared in load_extras rather than in the init method of BaseTrainer, leading to this error if load_extras is not called beforehand. As a best practice, class attributes should only be declared within init.

This PR addresses the issue by adding ema initialization to the init function of the BaseTrainer class and includes a check to ensure ema is not None when loading the state dict in load_checkpoint.

This fix resolves the error in my testing.