Created by: seiriosPlus

Incremental Learning Supported: At current, Trainer will run save_model at the end of a train. But, when we run PaddlePaddle with distribution, there are two problems need to be solved: **

1. there are multi trainers, they will all save a model in param_path, but there is no need for that.
2. the parameter server will not run save_model, but it needs to load model at startup. **

The solution is same as #10376 (closed) The different with checkpoint is:

1. save_model must be called by manual.
2. save_model only save models variables, do not need to save other things.
3. save_model do not delete files be saved.