Created by: chenwhql

This PR do two things:

1. The dygraph DataLoader use ListTensorProvider to init LoDTensor by Place in set_sample_list_generator, which split the Tensor init position of set_sample(_list)_generator and set_batch_generator, I think that is not good design.
2. The tensor init on CPU is ok, then it will copy to gpu pinned memory, then copy to gpu. The time cost has no change.