* [Cherry-pick] Apply IOU to test_parallel_executor_seresnext_base_gpu (#43812)
1. Fix the conflict between #43812 and current release/2.3 branch
2. test_parallel_executor_seresnext_base_gpu failed on 2 P100 GPUs with `470.82` driver.

* refine manual seed

* fix ci problem

* fix unittests

* fix unittest

* set is_init_py=false in manual_seed

* fix unittest

* fix bernoulli_op

* fix(unittest): change random_seed to manual_seed

* 🐞fix(unittest): fix manual_seed

* trigger ci

* fix test_sentiment

* fix test_imperative_save_load

* fix test_uniform_random_op

* fix test_uniform_random_op

* fix test_jit_save_load

* merge develop

* fix manual_seed

* fix manual_seed

* use global engine

* use shared_ptr

* fix double free

* fix bug

* fix bug

* fix bug

* fix test bug

* fix test bug

* fix test bug

* fix ci

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

* deprecate python memory optimize, test=develop

* remove memory_optimize in unittests, test=develop

* add unittests to deprecated interfaces, test=develop

* support sparse gradients
test=develop

* add unit test for fuse_opt_ops
test=develop

* 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

* add ir memory optimize test case for IfElse op, test=develop

* fix some unitttest failure by force using the python memory_optimize, test=develop

* tweak comments, test=develop

* fix unittest, test=develop

* fix unittest, test=develop

* add ir memory optimize test case for IfElse op, test=develop

* fix some unitttest failure by force using the python memory_optimize, test=develop

* tweak comments, test=develop

* fix unittest, test=develop

* fix unittest, test=develop

* Add Preface

* Add demo code

* Save file

* Refine code

* seems can work

* use elementwise strategy

* Use ElementwiseComputeEx

* Add comments

* extract functions from operator

* Refine code

* Follow comment

* code refine

* add op_fuse  pass

* add backward

* code refine

* use TopologySortOperations

* follow comments

* refine IsFusible

* code enhance

* fix op_fusion_pass

* refine code

* refine fuse_elemwise_act_op

* adjust the input and output

* refine logic

* add intermediate_edge

* disable inplace

* follow comments

* refine logic

* follow comments

* Remove the removable IntermediateOut

* change strategy

* code refine

* enable fuse backward

* code refine

* code refine

* rename unit test

* follow comments