* Imperative deep-first backward process (#16605)

* Fix bug of gradient interface

* shrink transformer

* Right transformer

* Change from width-first backward to deep-first backward process

test=develop

* Reverse iterator op's input

test=develop

* Polish code

* Change the iteration direction in ingrads' map slots

test=develop

* Polish code

test=develop

* test=develop, cherry-pick fix for transformer in dygraph

* test=develop, fix transformer in dygraph
/

test=release/1.4

test=release/1.4

test=release/1.4

* fix the bug of all_reduce_deps_pass
test=release/1.4

fix load bug about context.place

(cherry picked from commit bddb2cd3)

* fix preprocess script with processbar, integrity check and logs

* delete unnecessary empty lines, change function name

test=release/1.4

(cherry picked from commit bddb2cd3)

* Add linear learning warmup method

This warmup lr can be combinated with other learning rate strategies.
For example:
            decayed_lr = fluid.layers.linear_lr_warmup(
                fluid.layers.piecewise_decay(boundaries, lr_steps),
                warmup_steps, start_lr, end_lr)

* fix cudnn batch norm accuracy test=develop

* fix cudnn batch norm accuracy test=develop

* disable failed test for later fix test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

* fix compiled
test=develop

* follow comments test=develop

test=develop

* Add DecoupledWeightDecay

test=develop

test=develop

test=develop

* fix comments of 16410, test=develop

* modify inplace_op_inference_test according to pass interface change, test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop