* fix params with only 1 dim
* test=develop

checkpoint for distributed training.

* fix io doc
test=develop

* refine data_feeder doc
test=develop

test=develop

*  fix mkdir conflict

*  fix load/save lookup tables

 test=develop

* add lookup_table_utils

* fix load optimize vars on pserver

* delete lookup table utils

* fix save and load lookup tables

* fix load optimizer var

* fix load optimizer var, test=develop

* fix python 3 style, test=develop

* move lookup_table_utils to contrib utils

*  fix mkdir conflict

test=develop

* fix dim0 in _load_slice_up_vars

* fix dim0 in _load_slice_up_vars, fix innershape in delete_var_op

* Revert "fix lookuptable in reduce strategy"

This reverts commit 0e722c5e

* add unit test for dist

* add unit test for dist, test=develop

* cancel revert, test=develop

In the future, we'd like to encourage user to save
everything during training.
This allows us to
1. Do more flexible optimization passes
2. Re-train and fune-tune

* hide program APIs

* fix merge error

* update

add export_for_deployment flag to save_inference_model

Fix str and bytes problem
Fix sorted problem
Fix math problem
Fix CI problem