ENH: unify PADDLE_ENFORCE for cublas, cudnn, curand

update tensor usage in sgd-op

cmake: fix problem that go_library is never rebuilt.

* a simplest SGD op

Refactorize Tensor to Eigen convesion

add crop layer

Feature/add op test

Skeleton Of fully connected operator

2. Add crop layer unitest
3. Fix bugs

We put forward Op's inputs, outputs and output gradients into Grad
Op's inputs, and put forward Op's input gradients into Grad Op's output.
So Grad Op's `in_out_idx`, `input_format` and 'output format' need to be
rebuilt during Op creating.

Tensor should use PADDLE_ONLY_CPU macro