* sequential reader stage 1, test=develop

* fix ut, test=develop

* fix iterable=False reset bug, add some logs and polish code, test=develop

* inference feed partial data, test=develop

* Turn on keep_order=True for test, test=develop

* enhance ut to test more cases, test=develop

* test commit for reverting

* Revert "test commit for reverting", test=develop

This reverts commit 80aef42e.

* add ut of merged and unmerged results, test=develop

* add more uts for coverages and add en doc of api, test=develop

* follow comments, test=develop

* change note style, test=develop

* fix warnings oof gcc 8 compilation, test=develop

* fix boost::bad_get, test=develop

* refine PADDLE_ENFORCE, test=develop

* Enable fused_all_reduce_op_handle support GPU and CPU Gradients

* support sparse gradients
test=develop

* fuse optimizer

* fuse all_reduce
test=develop

* add fuse_parameter_groups_size
test=develop

* Polish code
test=develop

* Fix travis-ci
test=develop

* Add SetGroupAccordingToLayers and SetGroupAccordingToGroupSize
test=develop

* Add SetGroupAccordingToMemorySize
test=develop

* fix multi_devices_graph
test=develop

* reset params_grads
test=develop

* Polish code
test=develop

test=develop

test=develop

test=develop

Reduce one level of inheritence.

* Use Builder Chain to decorate new builders. It is easy to extend
  builders.
* Make graphviz path as a build strategy, not a FLAGS.

All leaves must be variables. When all variables are ready, the
execution will be completed. If a operator has no output, the `Op::Run`
might not be started when the execution of graph has been complete.