[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

* 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。

* Add casting initializers for bf16 training

* Changes after review

* Correct test and add comment

fix test sync_with_cpp (#32212)

* update, test=develop

* 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.

* add alias for fluid.contrib.mixed_precision

* add the weight decay func for the momentum op

* Add the multi_precision function in Momentum Optimizer.

* Make sure that the initial value of master weights are same with the fp16 weights.

* add static loss scaling.

* add the rescale_grad function in the pure fp16 training.

* use the original momentum updating method.

* Polish some codes, such as variable names.

* add docstring for apis.

* update the var creation details of _create_master_weight.

* not modify codes about imperative momentum updating.

* Fix the error of test_dist_sparse_tensor_load_momentum UT.

* add unit test for multi precision fp16 training.

* add more unit tests for CI.

* Use lower threshold values for allclose comparing in test_multi_precision_fp16_train UT.

* For CI Coverage Checking.

* add fp16 for layer_norm op

* revert layernorm api

* fix forward

* fix forward

* fix backward for layernorm with fp16

* fix unit test for layernorm with fp16

* fix with_mkldnn compile error for layernorm with fp16

* 1. revert to PADDLE_ENFORCE_NOT_NULL, 2. change static_cast<float> to static_cast<U>

* fix with_mkldnn compile error for layernorm with fp16

* fix with_mkldnn compile error for layernorm with fp16
Co-authored-by: Nzhiqiu <chenqiuliang@baidu.com>

* skip reader op in mixed_precision decorator

* add ut

* add fused_bn_add_relu op

Update amp_check_finite_and_scale_op and add an updating_loss_scaling op for static graph amp training. (#26240)

* update amp_check_finite_and_scale_op for static_amp.

* use amp_check_finite_and_scale in static graph amp.

* update grads to zero when grads own infinite values(as for amp_checkout_finite_and_scale op).

* add update_loss_scaling op in cpp.

* add update_loss_scaling_op unit test.

* update the doc of the check_finite_and_unscale op

* Update the process of gradients updating skipping if the gradients have infinite values.

* update the way to zero grads.

* update test_update_loss_scaling_op.py

* add log info when find infinite grads.

* add the unit test for UpdateLossScaling Layer.

* allow amp and recompute working together

* fix some typos in AMP. test=develop

* delete useless codes. test=develop

* add custom black varname test=develop

* fix dtype test=develop

* fix num test=develop

* fix ut test=develop

* fix coverage test=develop

* fix blackvar names test=develop

Optimize amp for multi-gpu to enable FP16 gradients transfer across gpus

init new amp, optimize inserting cast op for batchnorm

Change backward_guard to optimize_guard to maximize the allreduce overlap

test=develop

test=develop

* init auto loss scaling

test=develop

* change API.spec

* change ifelse to switch and use reduce_sum to optimize checking isfinite

test=develop

* Remove redundant code

test=develop

* Init mixed precision training interface

* Add fp16 test script

test=develop

* All initializers support float16

test=develop

* Code cleanup & add more code annotations

test=develop

* Update API spec

test=develop

* Add usage example in doc

test=develop