[cherry-pick 2.3] Add fused_multi_transformer op to optimize transformer generation performance (#42311)

* Add fused_multi_transformer op to optimize transformer generation performance (#41814)

* fix fused_multi_transformer compile failed in cuda arch < sm53 (#42315)

* fix ci timeout

add mixed-precission support for ipu

cherry-pick from #41733

* add DistributedFusedLamb op

* polish code

* fix compile error

* compatible with pten changement

* fix rocm compile error

* improve converage

* update upstream/develop

* fix cast_with_ptr.h

* add FLAGS_distributed_lamb_divide_nranks_when_allreduce=1

* fix clip before allreduce

* add use_master_param_norm

* code polish

* fix bug

* fix ROCM ci

* amp list updated

* tests updated

* gray list updated

* amp list updated

* test updated

* base changes for mul reimplementation

* empty commit

* tmp save

* full implementation of mul bf16/fp32 fwd bwd

* CI fix

* CI rerun

* changed unity build cmake to avoid gpu issues

* removed mul mkldnn from unity build

* added skipping tests if not cpu_bf16

* CI fix

* CI fix

* CI fix

* Fix scatter_op fp16 perf problem.

* Add scatter into black list.

* Add scatter into black list for dygraph.

* support FP16 for more ops

* add amp list tests

* refine reduce_mean_grad

* fix OP benchmark ci

* fix fp16 reduce_mean

* updat ut, but still have some problems

* remove mean/reduce_mean fp16 kernel

* support multi precision update for LAMB

* hide some api

* fix ci uts

* fix lamb output of dygraph

* remove some changes to some PR

* try to fix Py3 CI compile error

* fix test_imperative_optimizer, add lars ut, add layer_norm ut

* fix ut, fix format

* fix ut

* fix windows ci

本PR是fused_transformer的layer层代码，包含FusedFeedForward的layer层代码和FusedTransformerEncoderLayer的代码。

* Create stateful OneDNNAXPYHandler object.

This makes it possible to call it multiple times without recreating the
oneDNN primitives every time.

* Prepare SGDOpKernel to reuse its implementation from OneDNN kernel.

* OneDNN SGD kernel.

* Update call to use new OneDNNAXPYHandler object api.

* Setup seed in proper place.

* Enable OneDNN kernel only for single case.

* For dense param and sparse grad.

* Small refactor.

* Enable oneDNN by op attr or by cmd line flag.

* Use int64_t type for number of elements.

* Support dense param and grad from OneDNN kernel.

* Enable SGD OneDNN kernel when use MP BF16 optimizer.

* Force non-copyable/movable OneDNNAXPYHandler.

* Reuse OneDNNAXPYHandler for spare tensors in SUM op.

* Fix SFINAE rules.

* Remove recording event inside AXPY.

* Get rid of internal primitive caching.

* Stop use PP cache mechanims to store mem and primitive obj.
* Handler obj store and reuse needed desc & prim

* Do not derive from MKLDNNHandlerT

* Small corrections.

* Fix lr for bf16.

* Revert some changes.

* copy found_inf to cpu in advance to improve performance

* add npu test

* add npu test

* refine code

* refine memcpy op

* fix adam

* pass found_inf to adam

* add unittest

* fix bug

* refine unittest

* change unit test's directory

* disable unittest on cpu

* add clear_float_status op

* refine infershape

* fix typo

* refine check_finite_and_scale

* refine code

* reduce sum op default fp32, add into amp black list

* reduce_sum default fp32 can avoid return inf when the sum value large than 65504

* Add casting initializers for bf16 training

* Changes after review

* Correct test and add comment

* refactor_check_finite_and_scale_npu_kernel

* fix compile

* add alloc_float_status op

* add alloc_float_status op

* add FloatStatus for check_finite_and_unscale

* refine code

* remove unneccessary logic

* refine for fleet

fix test sync_with_cpp (#32212)

* Use the runtime to create the unsupported_fp16_list using in AMP.

* Add more infos about supported ops.

* Add some comments for the function of OpSupportedInfos.

* Fix the unit test of test_multi_precision_fp16_train.

* update, test=develop

* add fleet amp.init()

* add unittest for fleet_amp_init

* add cast ops before and after unsupported fp16 ops.

* Keep partial net in FP32 pattern.

* Support check_finite_and_unscale and update_loss_scaling for FP16 calculation mode.

* Add fp16 support for adam op.

* add multi precision attr for adam.

* Fix the bug of test_multi_precision_fp16_train UT.

* Code format for CI.

* Fix the redefine error about MPTypeTrait on windows.

* fix bugs of the _create_accumulators func in Momentum.

* fix bug when inserting post cast op.

* Add the update_loss_scaling op in allow_set of UnusedVarCheck.

* Update for ci coverage.

* Add some doc for OptimizerWithMixedPrecision.

* Fix the code style.

* Imporve the doc of `amp_init`.

* Change for fp16 testing if users have the infer program defined in separate way.