Fix a program copy regression.

Single device on se-resnet reduce from 0.56 to 0.50

paddle/fluid/framework/executor.cc

@@ -44,7 +44,7 @@ struct ExecutorPrepareContext {
   ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id)
       : prog_(prog), block_id_(block_id) {}
 
-  framework::ProgramDesc prog_;
+  const framework::ProgramDesc& prog_;
   size_t block_id_;
   std::vector<std::unique_ptr<OperatorBase>> ops_;
 };