Polish CompileTime InferShape

* conv_shift_op: initial implementation using Eigen

Limitations:
- both gradient outputs must be specified and are always computed
- explicit for loops => could be optimized in various ways
  (e.g., different memory layout)

* conv shift - gradient fixes

fix case when not all output gradients desired

* conv shift: minor cleanup

* conv shift - more minor cleanup

* conv shift: clean up & initial GPU implementation

* fix rebase issue

Init Python API

Fix OpDesc bug

remove unused PADDLE_ONLY_CPU comments

fix InferShapeContextBase to InferShapeContext

Add max pool op (with index)

Removed unreached code

ELU activation

1. InferShape should be a method for `OpDescBind` not `Operator`, since
  `OpDescBind` is a compile-time concept.
2. Pre-create operators and store them into a map. Make InferShape
   faster

fix compiler error: ‘unique_ptr’ is not a member of ‘std’

Implementing the fill constant op for the executor

add interpolation op

Use PROTO_LITE when refactoring Paddle

Add unit test of attributes in backward

* Adding implementation for copying a vector to tensor
* Changing Tensor test to access gpu memory indirectly

It will significantly reduce binary size. It is useful for mobile
deployment.