Add CudnnHolder and use it in Conv and ConvTranspose op

Optimize fusion gru

cuda windows

* fix high level API(Inference) bug

* patch the unit tests

modify the timer

Fix bug of sampling_id op

* Improve and fix fake_quantize_op.

* windows inference fix

* windows inference fix

add lac infer test

cuda math port

fix mac compie

update 0.14.0 changes to develop

refine prelu

Fix Mac compile error

fix sparse grad merge on pserver

delete pics from paddle repo and fix related docs

* remove template operator

* "fix compile"

* "fix ci"

* "fix ci"

add unit-test for chinese_ner

Fix generator not closed when iteration end bug