* add gradient Merge optimizer to meta, test=develop

Integrated Trainer of Parameter Server (API add `fluid.contrib.layers.sparse_embedding` only) (#22957)

* Integrated Trainer of Parameter Server

* Remove paddle.incubate.hapi.loss and reuse the paddle.nn.layer.loss in high level API

[Dy2Stat-ErrorMessage]  Add interface:create_origin_info_map and attach_origin_info for AST node (#25627)

* Add interface:create_origin_info_map and attach_origin_info for AST node. test=develop

* Fix code according to comments from reviewers. test=develop

* fix double grad visitid unit; test=develop

* change name hash_pair to HashPair; test=develop

* follow comment; test=develop

* reduce hidden size and loop number; test=develop

* change loop number; remove useless code; test=develop

Fix the argsort and sort op for the api2.0, and update the api 

* Remove the output for moving_average_abs_max_scale op, test=develop

* refine strategy compiler and meta optimizers
make async as a_sync

* unified signal error format

* refine signal error message

* remove ProgramTranslator.save_inference_model

* adapt save_quantized_model

* revert buffer check implemention

* remove useless import function

* simplify buffered reader to improve DataLoader performance

* fix 22 failed unittests

* fix cuda pinned context condition

* fix test_reader_reset failed

* fix two failed unittests

* change unittest place

* polish error messaage

* polish cast op GetExpecctedKernelType

* remove debug info in unittest

Enhance TracedLayer Error Message

Note: this PR uses assert to check type somewhere and check_type somewhere, the reason is that the check_type skips checking when it is under dygraph mode.

* saving inference model for user defined quantization model

* saving inference model for user defined quantization model

* fixed ci coverage

* add more settings for distributed strategy
Basically, DistributedStrategy has several parts of configurations:
- BuildStrategy: the same as paddle.fluid.BuildStrategy, but the distributed arguments are moved out of BuildStrategy
- ExecutionStrategy: the same as paddle.fluid.ExecutionStrategy
- collective communication configs: nccl_comm_num, hierarchical allreduce and so on
- distributed algorithms: async_update(mainly used in PS), lars, lamb and so on

* polish framework error message part3

* polish details

* fix error message print error

* Added DNNL cache management for DyGraph

* move FLAGS_use_mkldnn to more general CMakeLists, getu use of the flag in ClearGradients

* missing file

* Fixes after review

* Bringing back original idea of place for 'use_mkldnn' flag to be accessible from platform nad imperative.

* Removed duplicate and added docs

* Fixes for CI

test=develop

fix the cross OP behavior is not as expected when axis=0

test=develop

Based on the comment here https://github.com/PaddlePaddle/Paddle/blob/b5f8784cab94eae785659787fc529870c87b254c/paddle/fluid/framework/details/build_strategy.h#L49

The unit test which compares Reduce and AllReduce must have diff. The PR_CI_Night runs on P40 machine and it has 8GB GPU, which is smaller than the 16GB normal CI machines. So we decrease the batch size in the past to make it runnable: https://github.com/PaddlePaddle/Paddle/pull/24651/files  . Decreasing the batch size makes the difference occurs often. So this PR replace the absolute delta by relative delta.

Before this PR, the unit test failure happens with probability about 1/100. After this PR it doesn't happen.

* fix the double grad bug for the star gan. test=develop

* update the retain_graph parameter doc. test=develop

* add the unit test for the retain_graph parameter. test=develop