[Feature Request] Update the energy bias after the final step of training

[njzjz]

Summary

Add an option to update the energy bias after the final step of training.

Detailed Description

Sometimes, the predicted energy is shifted compared to the reference energy, such as #2935. I think it might be because the energy loss is too small (the force loss cannot train the energy bias). However, we have an easy way to improve it: do update_energy_bias after the final step of the training. It just changes the energy bias and will not break other things. We already have the code for update_energy_bias, so we can add an option to call it.
(in theory, it is equivalent to --finetune with 0 steps, but a user won't come up with it.)

Further Information, Files, and Links

No response

[wanghan-iapcm]

I agree. we can provide the feature after @iProzd refactorizes and merges the update_energy_bias with the current output stat code.