Add 4 non-fault-tolerant demos

[helinwang]

Related: #147

[Yancey1989]

As this implement for reader, each trainer will fetch the same training data, maybe the trainer would fetch a part of the whole training data, or i missed something?

[typhoonzero]

In non fault tolerant mode, readers must fetch part of the training data by it self.

[helinwang]

Thanks! Sharding does not work correctly before #319 and #318 is fixed, can I merge this first and submit a follow up PR for sharding after these issues being resloved?

The planned code is:

def get_shard(paths):
    trainer_count = int(os.getenv("PADDLE_INIT_NUM_GRADIENT_SERVERS"))
    files = glob.glob(paths)
    files.sort()
    flie_count = len(files)
    files_current_train = []
    for idx, fn in enumerate(files):
         if idx % trainer_count == trainer_id:
             files_current_train.append(fn)                                                                                                                                                                
    return files_current_train

[Yancey1989]

LGTM!