Using Smart pointer to optimizer memory usage of dyGraph

[JiabinYang]

This PR use smart pointer to hold grads and release the fwd vars. In this way, we can save lots of memory:

Test on Resnet, with Tesla K40m(Mem 11439MiB):

Without smart pointer : max batch size -> 56
With smart pointer: max batch size -> 99

And We test the convergence of several models under models/dygraph the result is as follow:

Mnist:

ptbRNN:

reinforcement learning:

transformer:

[JiabinYang]

This Commit will enhance the performance by 15-20%:

The test result is based on Tesla K40m

PTB_model:
before: 82.2777 s per 1/10 of all data
after: 62.12738 s per 1/10 of all data

Resnet. Batch size 50
before: 1.49s per batch
after: 1.26s per batch

[sneaxiy]

LGTM.