Remove memory copy from GPU to CPU in the linear chain crf op.

[lcy-seso]

At the time when implementing linear chain crf operator, the framework does not support operator running on different devices. The operator has to copy data it needs itself from different devices.
https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/linear_chain_crf_op.h#L68

But, not this problem is fixed now. So memory copy from GPU to CPU in the linear chain crf op should be removed.