* "derived cudnnDevice context"

* "leave remove cudnn handle from CUDADeviceContext"

* "fix math function error"

Also add `elemwise_sub/mul/abs/clip` layers

* Add fill_op

* Fix bug

There are mainly following fixes:

- take `DeviceContext` as the template parameter of math functors and OpKernel instead of `Place`
- remove `eigen_device` interface in base class  `DeviceContext`
- remove `GetEigenDevice` interface in `ExecutionContext` and base class `DeviceContext`
- remove unused `platform::EigenDeviceConverter`
- rename `REGISTER_OP_GPU_KERNEL` to `REGISTER_OP_CUDA_KERNEL`
- rename `USE_GPU_ONLY_OP` to `USE_CUDA_ONLY_OP`

* Updating the Latex equation for Adagrad

* Fixing Latex euqations for adadelta, adam and adamax

* Fix documentation

* Address review comments

* Fix compiling error of gcc4.9.

* Refine the check of cxx compiler flags in api/CMakeLists.txt.

* init decoder_trainer

* can run

* fix lod

* add sharelod to cross_entropy_grad_op

* add avg_cost to fetch list

* modify learning rate

* can run

* optimie code

* add early exit

* fix print

* revert test_understand_sentiment_conv.py

* add act to fc

Add LRN efficient GPU implement

* Using latex fraction syntax in sigmoid and logsigmoid op

* Fixing the rendering of the latex equations in adamax operator