feature/parallel_gpu

* "init use all default devices"

* "fix init test"

* add lod tensor ToAbsOffset test

* add share lod to topk op and softmax op

Added backward.calc_gradient to backpropagate gradient from given targets to inputs.

* test dist word2vec

* multiple trainers work

* Add Seq2BatchFunctor, which will be used in WarpCTCOp.

* Implement WrapCTCFunctor and WrapCTCKernel.

* Add unittest of warpctc_op.

* Modify the check_output inferface in python unittest framework to allow check a subset of outputs.

* Use absolute offset lod in warpctc_op and related functors.

* Refine the comments of warpctc_op.

* The new python unittest supports checking a subset of the outputs, so revoke the previous change.

* Rename the transform from LoDTensor to Tensor with shape [max_sequence_length, num_sequences, sequence_width] to PaddingSequenceFunctor.

* Update to the newest codes.

* Rename the PaddingSequenceFunctor to PaddingLoDTensorFunctor and remove the computation of dimensions out of the functos.

* Rename Tensor::CopyFrom to Tensor::Copy

* Fix CI

* Fix compile

* Remove unused LoDTensor methods

* Update

* create tensor in recv op

* static global function to global function

* "reuse ShareLoD with no regret"

* "removed base class shareLayout"

* "fix CI"

* add rename guard

* add device_data_transform

* add device_data_transform_test

* modify GetExpectedKernelType

* update operator.run

* support test test_label_semantic_roles

* optimize code

* optimize code

* rename GetActualKernelType to GetExpectedKernelType

* fix chunk_eval_op and device_data_transform_test

* add is_same_place to place

* optimize code, refine rename_guard

* refine rename guard, add GetKernelTypeForVar

* optimize code

* add some log

* rename guard

* use sub scope to create var

* fix compile

* add IsInitialized for Tensor

* add VarIsTensor

* fix op_registry_test

* test

* tmp disable priority

* restore switch_kernel.md

* code clean

This can make it easier to locate error.

It will be used for LoD information in LoDTensor since LoD is a copy
on write field.

It is pretty slow for copying LoD information between operators. For
resnet it will cost roughly 10% time of whole time, including reading
data.