- The trainer and inferencer will load params from disk if param_path
  argument is not None in their constructor.

- Remove params.py, we will expose core.Scope to the user if needed
  (e.g., for GAN). Currently we will not expose it, unless we clearly
  know doing so can support GAN.

- Add `save_params` to Trainer (a TODO item).

- rename "network" to "program"

A naive implement trainer.train by executor

add some instructions for running vgg distributedly

Fix cpplint format check

* "fix double type error"

* "fix ci"

Use multi thread to do update

add lookup_sparse_table_op

* Fix cpplint issues in sequence_padding

* Fix typo in cu file

* Fix dependencies of sequence_padding

* Add include

* Fix CPPLint issues in fluid/inference/tensorrt/

* Fix compile errors

Fix unittests

install ccache in Dockerfile to speed up compile

Fix shfl_sync for CUDA8.0

Add float16 support to save op

Change `customize_loss_grad` to `use_default_grad_scale`.

Fluid new API scaffolding

* Fix CPPLint issues in lstm_cpu_kernel.h

* Fix CPPLint issues in math/math_function_test

* Fix CPPLint issues in math/math_function_test

* Fix CPPLint issues in math/concat.cc

* Fix CPPLint issues in math/concat.cc

* Fix CPPLint issues in math/concat.cc

* Fix CPPLint issues in math/gru_cpu_kernel

* Fix CPPLint issues in math/selected_rows_functor_test.cu

* Fix compile error

* Fix compile error