cherry-pick #22615

* Add support for dynamic_decode(while) training. test=develop

* Fix assign_op and tensor_array_read_write_op after solving conflict. test=develop

* Fix test_rnn_decode_api.py. test=develop

* Refine docs for apis in rnn.py. test=develop

* Adjust outputs of dynamic_decode. test=develop

* Remove the force_cpu update in assign_op. test=develop

* Remove the force_cpu update in assign_op. test=develop

* Make RNNCell.get_initial_states support batch_dim_idx argument. test=develop

* Rename _create_array_outof_while as _create_array_out_of_while in rnn.py.

test=release/1.7

This PR makes assign op support LoDTensorArray and enable the loop_vars in
while_loop to support tuple or list.

* test=develop, bug fix for sparse recorder

* update layers, test=develop

* update layers for resnet, test=develop

* fix is_test attr, test=develop

* update cycle_gan, test=develop

* update reinforcement_learning, test=develop

* update ocr, test=develop

* fix bug, test=develop

* add example for reshape op

* modify fc to linear in sample code, test=develop

* remove FC, test=develop

* remove warnings, test=develop

* drop fluid/imperative/README.md , test=develop

* change fc to linear, test=develop

* polish code style, test=develop

1. Add a new input named batch_roi_nums for prroi_pool_op. batch_roi_nums includes the number of roi for each image in batch when rois is Tensor. This information is saved in rois's lod when rois is LoDTensor.
2. add grad check to prroi_pool_op and solve unnormal X grad diff in CPU.

* add erf op and python interface.

* add fp16 support for erf op.

* add unitests for erf op and its python interface.

* fix dim=[] bug in reduce ops

* append optimize op in the grad block of current block if current block is in control flow. test=develop

* add conditional grad op when optimizer used in control flow. test=develop

* add comment and modify typo. test=develop

* fix append_backward to support control flow. test=develop

* add test. test=develop

* fix copy_var_to_parent_block and conditional_block_grad. test=develop

* fix bug: revert to append conditional_block_grad vars to sub grad block. test=develop

* fix bug: revert to assign var to parent block even if var already is in parent block

* fix bug: consider outputs is empty. test=develop

* move _rename_grad_ out. test=develop

* modify code according to reviews from Huihuang. test=develop

* modify code according to reviews from Jinle. test=develop

* add shape description in accuracy api doc test=develop

* fix details test=develop

* fix details test=develop

* Fix example in log_loss doc, test=develop

* Use none for unknown dim, test=develop, test=document_fix

* convert dtype of vars for range op, test=develop

* update error information about batch_norm_grad

* update bn,test=develop

 fix the doc of scatter

* update layers used in mnist dygraph model, test=develop

* fix import issue, test=develop

* add dygraph utils, test=develop

* add unittest, test=develop

* fix stat shape back in global auc scenario
* add UT to cover global auc

* update layers, test=develop

* fix input numpy, test=develop

* fix bugs, test=develop

* follow commments, test=develop

* update getitem, test=develop

Update iou_similarity op to support non-normalized bbox

* add unittests, test=develop

* set dtype of compare op to bool, test=develop

adds unary operator __neg__ for VarBase in dygraph mode, and for Variable in static graph mode.

* increase example code of py_func, fix some wrong description of English API doc

* yolo_box OP add Attr(clip_bbox). test=develop

* patch math method for varbase using auto-generated op functions, test=develop

* clean code that handles batch_size, test=develop

* follow comments, test=develop

* follow comments, test=develop

* code clean, test=develop

As the title

Add basic while_loop

Add tests to use dy/dx to make sure the gradient values calculated by the control flow backward is correct. Also fixed bugs detected by those tests.

Fix bugs:

1. Unlike sum_op, optimizer ops don't allow uninitialized input tensor. But in conditional_block_grad_op, since the conditional_block may not run, the output gradient tensor may be uninitialized, which will cause the optimizer op error. To fix it, we should let optimizer ops support uninitialized input like sum_op or assign the uninitialized gradient to 0 when the conditional_block_grad_op doesn't run. I found there are about 10+ optimizer ops. **To be simpler, I just assign output gradient of the conditional_block_grad_op to 0 in this PR**. But it can be further explored whether we can make optimizer ops like sum_op to support uninitialized input tensor because theoretically we can speed up without the assigning in conditional_block_grad_op.

2. Infer parameter shapes during append_backward. I didn't know that all our parameters are in global block. When op_desc is inferring shapes at the sub-block, it may not know the shape of gradients of parameters whose shape information is at global block. I fixed it by inferring shapes of gradients from forward var.

This PR also did some code clean up:
1. Print the var name when sgd_op catches shape error so that it is easier to debug
2. Fix a typo: dicta -> dict

    Add a python interface for Gelu.
    Add documentation for fluid.layers.gelu.