Created by: zhangting2020

fix the bug:

• when GpuMemcpy is included in compute, it will be counted repeatedly in the compute ratio:

recurrent_grad/sum/compute                                                 2640        83.2866     62.121459 (0.745876)    21.165112 (0.254124)    0.015171    0.095683    0.0315479   0.068406
  recurrent_grad/sum/compute/GpuMemcpyAsync:CPU->GPU                       570         8.39523     7.405052 (0.882055)     0.990173 (0.117945)     0.010849    0.046229    0.0147285   0.00689527

This will cause the ratio of compute time to be higher than the actual value.

• GpuMemcpy may be in main events, it should be counted.