Sync computation when Python invoke `run`

* Since GPU is an async device by default. We should sync computation
  when Python invoke `run`. So Python can get the correct computation
  result

paddle/platform/device_context.h

@@ -34,13 +34,14 @@ class DeviceContext {
 
   template <typename DeviceType>
   DeviceType* get_eigen_device() const;
+
+  virtual void Wait() const {}
 };
 
 class CPUDeviceContext : public DeviceContext {
  public:
   CPUDeviceContext();
   explicit CPUDeviceContext(CPUPlace place);
-  virtual ~CPUDeviceContext() {}
 
   Eigen::DefaultDevice* eigen_device() const;
 
@@ -59,7 +60,7 @@ class CUDADeviceContext : public DeviceContext {
   virtual ~CUDADeviceContext();
 
   /*! \brief  Wait for all operations completion in the stream. */
-  void Wait() const;
+  void Wait() const override;
 
   /*! \brief  Return place in the device context. */
   Place GetPlace() const override;

paddle/pybind/pybind.cc

@@ -238,7 +238,13 @@ All parameter, weight, gradient are variables in Paddle.
              return Backward(forwardOp, no_grad_vars).release();
            })
       .def("infer_shape", &OperatorBase::InferShape)
-      .def("run", &OperatorBase::Run)
+      .def("run",
+           [](OperatorBase &self,
+              const Scope &scope,
+              const platform::DeviceContext &dev_ctx) {
+             self.Run(scope, dev_ctx);
+             dev_ctx.Wait();
+           })
       .def("type",
            [](const OperatorBase &op) -> std::string { return op.Type(); })
       .def("outputs",