First half translation of operator definition.

Sync computation when Python invoke `run`

Simplify GetAttrType code

Stabilize prelu gradient check

lstm unit

enable VGG with MKLDNN layers

Add gemm with stride arguments

add namespace and reduce warning

Update annotation in layers.py and update FAQ.

Add multiplex operator

implement framework of seq_pool_op and its unitest

adding doc for comparing cluster bootstrapping tools

Change LoDTensorDesc::dims from int32 to int64

make RecurrentOp's backward compatible with framework

Fix progma_once error

Change namespace of pybind.cc to pybind

* Since GPU is an async device by default. We should sync computation
  when Python invoke `run`. So Python can get the correct computation
  result

Remove `numel` field in tensor

* It is duplicated with `dim_`. We can use `dim_` to calculate `numel`
  everytime. It does not cost too much.
* `numel` is not initialized by constructor. Also, `numel` is hard to
  synchronize with `dim_`.

So just remove it.