[dy2static]implement tensor.cuda() in static graph

[liudongxue01]

PR types

Others

PR changes

APIs

Description

• tensor.cuda() is empty implementation in static graph mode, but on the dygraph mode, this api performs the memory-copying behaviour(H2D or D2D).
• So on dy2static mode, the model result would not be equivalent as on dygraph mode.
• Upon above, we implemented the memory-copying logic code in this api using 'memcpy' operator.
• We alse add some fixup logic in interpreter_util.cc as a workaround of the bugs in 'memcpy' kernel.
• Note: 'memcpy' operator does not have corresponding backward node, so we set 'stop_gradient=True' on the result tensor object.

Others

Pcard-73789

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[sunzhongkai588]

LGTM for docs

[From00]

LGTM

[jeff41404]

LGTM