* fix stop_gradient in paddle.to_tensor

* make to_tensor copy if data is varbase

* add ut

* refine sample code