* add start_record interface

* call master client in reader

* update

* add demo code in comments

* update comments

* delete unittest for recordio reader

* fix recordio split and task passes

* update for pre commit

* update

* update, still need to sync client wait for pass end.

* able to sync passes for task dispatching

* update to comment

* update

* fix yapf check

* why local pre-commit fails? version is the same

* fix race condition

* update

* fix race condition

* this still have duplicate problem in unit test

* update

* update

* update by comment

* update

* default learning rate, temperary fix

* update

It's already fixed by Wuyi's PR, but his PR may take some time to
merge, but I want to get this change in ASAP.

Unchecked errors could be handled by: cd go; gometalinter --vendor --disable-all --enable errcheck $(glide nv)

* using etcd as fault tolerant training

* update

* workable version, ft not tested

* small fix

* update

* remove TODO

The buffer allocation size should be number of bytes, not number of
floats.

* Server recovery from checkpoint

* init etcd cclient

* add etcd

* add etcd.go

* fix compile problem

* move code to etcd.go

* add etcd_lister.go for pserver client

* add etcd_client_test.go

* merge etcd_client_test and client_test

* refine client_test.go

* refine code

* format code

* add TODO and use interface instead of struct

* fix typo of initDesiredPservers

* optimize dir structure of go/pserver/client

* add a flag to config index for pserver

* follow comment

* fix path

* optimize code

* remove err in pserver NewEtcd

* restore comment about /ps_desired