* Support memory eager deletion on recurrent OP (#17710)

Test PaddingRNN on V100 GPU device.

Test configuration: large model, padding mode (which is the mode using recurrentOp), one GPU.
                   
GPU memory (MiB):   6414 (this PR)     vs   6837 (without this PR)
Speed (steps/s):         10.28 (this PR)    vs    9.89 (without this PR)

* Fix random test_recurrent_op failure (#18718)

The change includes 3 things:

1. Set CPU_NUM to 1 in the tests because the ParallelExecutor will print warning that CPU_NUM is not set and use default 1.

2. Old tests compare two RNNs, hand written simple RNN and same RNN built by Paddle, but initialized RNN weights in numpy random and Paddle random separately. Fixed it by setting weights and bias values.

3. Also set numpy random seed in the tests. Now the two RNNs diff can be smaller (rtol from 0.1, 0.2 to. 0.01) in the tests.

* update executor feed

This PR cherry-picked the fix of multiple Layers parameter missing error in dygraph mode，the original one is #18968

* fix bug in Class MultiSlotDataGenerator's function _gen_str, test=develop (#18222)
* fix some bug when merge sparse embedding parameters, test=develop (#18223)
* fix communicator with pyreader (#18350)
* delete AllocatorFacade destructor  (#18606)
* fix distribute transpiler GRPC error code 4, RPC Deadline (#18984)
* merge pr #18441

* fix optimizer bug
test=develop

test=release/1.5, fix problem that get_attr method can't using default mode when we call has_attr in dygraph (#19328) (#19414)

* add default getItem

* test=develop, fix has_attr disabled error in Layer

* test=develop, fix GroupNorm and deepcf bug on attrs

* fix depthwise conv gpu kernel bug, test=develop
* add more depthwise conv test, test=develop

* make_roi_perspective_transform_op_return_mask_and_matrix

* make_roi_perspective_transform_op_return_mask_and_matrix

* CHERRY_PICK 18941, 18860: TRT fp16 support.

test=release/1.5

* CHERRY_PICK 19213: Fix BUG: Mask RCNN inference diff When using AnalysisPredictor.
    1. fix affine channel fuse pass.
    2. fix condition block op.
    3. fix merge lod tensor op bug.
    4. fix memory optim cause by reset lod op.

    test=release/1.5

* add fusion_seqpool_cvm_concat test=develop

* simplify pass, test=develop

* fix code style, test=develop

* Add call stack info during runtime and compile time
test=develop

* add warning info for CPU_NUM
test=develop

* update dygraph parallel.py
test=develop

* prune the feed op in compiler
test=release/1.5

* remove compile from PE
test=develop

* test CUDAPinnedPlace in reader
test=release/1.5

* fix REGISTER_OP_WITHOUT_GRADIENT
test=develop

* fix backward bug

* test=release/1.5, cherry-pick hide not_support for dygraph

* test=release/1.5, cherry-pick hide not_support for dygraph

Add path to handle 1D vector

* Update lamb optimizer (#18333)

* Update lamb optimizer

* Regenerate api spec

test=release/1.5

* Give an experimental warning

test=release/1.5

* sequnce mask support max length tensor input; test=develop

* add rnn_impl.py; test=develop

* add basic gru lstm unittest; test=develop

* fix api spec; test=develop

* fix sequence_mask op bug;
test=develop
test=document_preview

* change +-*x to elmentwise_op; test=develop

* add mkl flag; test=develop

* fix rnn impl bug; test=develop

* update api spec; test=develop

* fix doc bug; test=develop

* fix lstm bugs; test=develop

* Revert "Cherry pick install check for multi gpu (#18245)"

This reverts commit d0219002.

* test=release/1.5, ci start

* fix slice op bug; test=develop

* fix variabel test bug; test=develop

* remove slice while true; test=develop

* test=develop, add add_multi_gpu_install_check (#18157)

* test=develop, add add_multi_gpu_install_check

* test=develop, refine warning doc

* test=develop, refine warning doc

* test=develop, refine warning doc

* test=develop, support multi cpu

* test=release/1.5, cherry-picked from develop

* Update backward appending stragety to support double backward and fix some bug. (#18104)

* Update backward.py:
     - If there is no input grad var in all outputs of previous ops, do not append this op into graph.
     - Only apply this stragety when double backward.
* Update some double backward op.
* Update sum_op to judge whether a tensor is empty by numel or IsInitialized().

* fix save/load in fleet (#17675)

* fix save/load in Fleet
* add UT framework of Fleet (#18058)

* add paddle cloud role maker for customized usage, note this is only for industrial users that have cloud environment pre-configuration (#18121)

add paddle cloud role maker for specific cloud usage. This pr will simplifies user's configuration in distributed training.

* assign role_maker before use (#18137)

Cherry pick retinanet_target_assign_op(#17893), sigmoid_focal_loss_op(#17895) and retinanet_detection_output_op(#17896) for supporting retinanet (#18141)

* test=release/1.5
Fix conflicts in test_layers.py when adding target assign operator for supporting retinanet. Cherry pick #17893

* test=release/1.5
Add sigmoid focal loss operator for supporting retinanet. Cherry pick #17895

* test=release/1.5
Add detection output operator for supporting retinanet. Cherry pick #17896

* test=release/1.5
fix wrong code style in test_layers.py when cherry pick retinanet_target_assign #17893

* test=release/1.5
Fix type error of std::pow in sigmoid_focal_loss. Cherry pick #17895

* fix logging unable. test=develop

* unset sys.stdout for stream handler. test=develop

* fix newly add basicConfig. test=develop

* fix import error. test=release/1.5

* remove nccl dep when the number of GPU is 1
test=develop

* use multi card run syncBN
test=release/1.5

Add cascade rcnn support.

cherry pick for deform roi pooling

* fix py_reader iterable bug, test=release/1.5

* move data from buffered_reader,test=release/1.5

* update CPU_NUM config
test=develop

test=release/1.5
cherry-pick from #17952
The scatter op has a calc bug when the indices has same index, the scatter op use overwrite mode to calculate the same index, fix this bug by using the accumulate mode to calculate the same index.At the same time, the gather op has the same bug when the op calc the grad. And we use the lib of open-blas and eigen to optimize the time cost in accumulate mode.

test=release/1.5

* Cherry-pick fix random Python3 CI failure.

In some tests, SWEs used "print('xxx').format('xxx')". The syntax
is only supported in Python2, not python3. However, since those
lines are related to data download, if the CI machines already have
the data, it passes CI tests. That causes random failure.

* Cherry-pick: disable CUDNN case of test_warpctc_op

test=release

TODO: it is a temporary fix for Paddle release 1.5. We have to fix
this failed unit test soon.

test=develop

* Enable seq_pool op to accept len 0 input

test=develop

* Update sequence_pool's api

test=develop

* Add more unittest cases for seq_pool op

test=develop

* Remove legacy comments

test=develop

* Don't use template in op maker

test=develop