Add Python Reader Op (Python side and unittests)

Distribute transpiler handle adam accumulator

fix bcast with parallel executor + cpu

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into distribute-transpiler-handle-adam-accumulator

Merge branch 'complete_py_reader_python' of https://github.com/sneaxiy/Paddle into complete_py_reader_python

Merge branch 'complete_py_reader_python' of https://github.com/sneaxiy/Paddle into complete_py_reader_python

Throw all warnings from git describe in setup.py.in

Fix optimizer accumulator

add unittests for Reader ResetAll

Merge branch 'fix-optimizer-accumulator' of ssh://github.com/jacquesqiao/Paddle into distribute-transpiler-handle-adam-accumulator

Ckpt new api

Complete hsigmoid_op

fix FLGAS_rpc_deadline does not effect

Merge branch 'complete_py_reader_python' of https://github.com/sneaxiy/Paddle into complete_py_reader_python

Merge branch 'complete_py_reader_python' of https://github.com/sneaxiy/Paddle into complete_py_reader_python

Merge branch 'complete_py_reader_python' of https://github.com/sneaxiy/Paddle into complete_py_reader_python

Merge branch 'complete_py_reader_python' of https://github.com/sneaxiy/Paddle into complete_py_reader_python

Merge branch 'complete_py_reader_python' of https://github.com/sneaxiy/Paddle into complete_py_reader_python

Make scope_buffered_ssa_graph_executor Exception safe

Add Unsqueeze operator and unit testing

Fix the HierarchicalSigmoidGradOpKernel and refine the codes. Now hsigmoid_op is same with V2 implementation and can pass gradient check.