repair image problem of en distributed training

1. add 1 sentence that was not translated in the original version
2. repair image showing problem

Add python wrapper for Adadelta optimizer

Enhance sequence_expand operator

Add float16 support to dropout operator

Fix bugs in c++ readers

add math_function to softmax's dep list

move distributed lookup table design to fluid/dist_train

Repair deadlink of fluid doc

repair link of "Executor"

Fix compilation for gcc5.4

* Support copy in Fluid channels

* Address PR review comments

Enhance LoDResetOp and add python wrapper

Disable the link flags on Mac.

Enable P2P memory copy

On k40 with 4 devices, time reduces from ~4.0 to ~3.8+, should be
more obvious on better hardware

move two images