[Feature] Test Both Global Model and Personalized Model

[GaoxiangLuo]

Is your feature request related to a problem? Please describe.
In the loop of the trainer, the evaluation is always after training. Hence, the model performance reported isn't from the pure global model but rather taking the global model and then conduct fine-tuning on the local dataset. Both models are meaningful, so we may want to report both.

Describe the solution you'd like
In the final round, conduct an evaluation before the training and call it the performance of the global model. Then, record the evaluation after the training and call it personalized model.

Describe alternatives you've considered
We can also save global model on aggregator registry and personalized model on trainer registry, and then conduct testing offline.