Enable gradient clipping.

[qingqing01]

This issue is related to #775 and #1891. I'll try to fix it. On the other hand, is it necessary to enable gradient and error clipping in gru-memory and lstm-memory which are implemented in one cpp file (not in the recurrent layer group)? @hedaoyuan @lcy-seso

[lcy-seso]

Is it necessary to implement norm gradient clipping in PaddlePaddle, for example, constraint a certain norm of gradient matrix by a threshold?

[lcy-seso]

Gradient clipping is critical to a stable training process of RNN models, this issue #1961 is also related to gradient clipping.