* Since GPU is an async device by default. We should sync computation
  when Python invoke `run`. So Python can get the correct computation
  result

Using DeviceContext, not Place to get stream

* Also use `const DeviceContext&` all the time, to prevent `const_cast`

Fix #4169
Fix #3468
Fix #3475

* refind accuracy_op

* follow comments

* follow comments

Note: If no symbol with a suitable value is found, both this field and dli_saddr shall be set to NULL.

Fix TravisCI

It is buggy in some Linux because the unique_ptr will be free however
the std::string trying to use that char*.

Moreover, it is no need to demangle for error log by Paddle.
Just use `c++filt` or other shell utilities to do this.

* with unit-tests
* Also complete `memcpy`

1. finish lookup table CPU and GPU kernel
2. Add some cuda helper
3. Add some math funtor

* Wrong Printf format before

* Should always use `dynload::` for cuda function.
* Fix cublas.h without DSO load.