Created by: zhiqiu

This PR try to split VarBase from python Variable Wrapper, thus Paddle can use VarBase which is exported from C++ directly in dygraph mode.
The reason to do that includes,

• to decouple Python side logic of dygraph eager execution with static graph,
• to speed up training process by directly accessing C++ VarBase,
• to cooperate with auto generated OP functions, which may be introduced in the future.

After this PR, in dygraph mode, all the APIs of Variable before are indeed of VarBase. For example,

• to_variable() returns a VarBase object which has the same interface with Variable,
• layers.xx accepts VarBase objects as inputs and returns VarBase objects after some operations,
• create_parameter returns a ParamBase object which is a child object of VarBase.

Models evaluation results. (using ptm with default params in models repo.)

• Conclusion:

  • in dygraph, 8% speed up and no effect on astringency;
  • in static graph, no effcet on speed and astringency.

• Dygraph, before this PR, time:208.4s, ppl:265.4
  

• Dygraph, After this PR, time:185.0, ppl:267.3

• Static Graph, before this PR, time:43.8s, ppl:268.1

• Static Graph, after this PR, time:44.1s, ppl:268.7