Cannot set different data related parameters for different clients.

[DavdGao]

As the title says, since the dataset is loaded before both fedrunner.run() and the setup of clients, the config_per_client.yaml doesn't work for the dataset, such as different batch size.

[DavdGao]

• Also, when FS finish the initialization of the global config, all unused configs are removed. However, if someone want to run a certain algorithm for some clients by the --cfg_client xxx.yaml, it will report attribute error since all related parameters within the config have been removed and the default value of cfg.${METHOD_NAME}.__new_allowed__ is False.
• Another bug is that when combine the global_cfg with the client specific configs, the global_cfg will call the assert function s again. Considering all unused parameters have been removed, some assert function will report attribute errors.

[xieyxclack]

We need a discussion about how to resolve this issue. @joneswong @yxdyc @rayrayraykk