Load saved embeddings with different numbers of trainers

[zheng-da]

Currently, we have to make sure the number of trainers is the same when we save and load embeddings. However, it is common that people change the number of trainers during training and inference. We need to modify our save and load functions to support this case.