To support sparse parameter update in distributed training

[wangkuiyi]

support distribute training in python v2 API:

1. support sparse remote updater.
2. support save parameter in trainer.

How to Test If We Made It

1. Start parameter servers:

   sh paddle/demo/quick_start/cluster/pserver.sh

2. Use example word2vec to start a distributed training job:

   In file paddle/demo/word2vec/api_train_v2.py, change cluster_train = False into cluster_train = True.

3. Start word2vec distributed training

   sh paddle/demo/word2vec/api_train_v2.py