1. rm Global.eval_mode
2. add Global.task
3. mv type_name to ppcls.utils
4. build dataloader, loss, metric by mode

1. ModelSaver();
2. _build_ema_model();
3. _init_checkpoints();
4. others.

1. rm engine.device and use paddle.device.get_device() instead;
2. mv some code to build_train_func or build_eval_func to simpfy engine.

https://github.com/PaddlePaddle/PaddleClas/pull/2633

evaluate once every 400 iters