enable inference benchmark

* Enable the case N != ldc in EigenBlasGemm.

* Use MemoryHandle instead of direct calling of posix_memalign to alloc temporary memory.

* Use Eigen's slice() instead of a temporary memory.

* Add if-else for different cases in EigenBlasGemm (for N ?= ldc).

Use protobuf v3.2.0 for MOBILE_INFERENCE compiling.

* Fix Proformance problem of enforce

* Fix missing `;` in code

* Fix CI

* add gflags for C++ unittest

* Translate the CPU profiling document

* Paragraphing

* Unify fluid submodules to fluid module

Change books just use `import fluid`, not submodules

* Remove g_main_program/g_startup_program

Use default_main_program/default_startup_program instead

* Typo

* Add API for switch default program

* Two functions: switch_main_program/switch_startup_program
* A guard: program_guard. Users can use the `with` statement change
  default programs
* Change unittests in `test_layers`

* Fix CI

* Fix CI

* Fix CI

Fix #6087

* fix python gflags init

* format code

Revise comments in rank_loss_op

* add back print_operators_doc

* fix style check

* fix style check

Amend license and comments in reshape_op

Refine doc for smooth l1 loss op.

* Add survey of support of half in different CUDA versions

* small fix

My unpool max 2d

fix conv and conv_trans op doc

* Make param_attr as a strong typed class

Fix #5819

Small fix in profiler.py of fluid.

Fix gru_op related code style