Enhance watcher behavior

[alvalin0203]

Expose grpc cancel error to etcd client to avoid client hang forever
Relative issue refer to #1227
Main modify in watch.py
Root cause：1. If there is no any interacting between client and server for specific time（for example 5minutes），that leads server delete client‘s token，which client is not aware of that，when client rebuild a new watch，grpc server would return response，which response's create property is true，but cancle property also is true，in original etcd code that just judge create property，would regard this response as successful which would lead to the client thought watch request is successful, but in actually failed, therefore the client would never received any response any more. So in this fix，expose server watch cancel response as canceled exception to client to make it rebuild token
2. Fix in watcher run function handle exception part which lose control of lock part， therefore would cause concurrence problem