* Added cuBlasLtHandle_t to device context.

* Added fused_gemm_epilogue op.

1. Added fused_gemm_epilogue op to leverage cuBlastLt Epilogue.
2. Support fusion Act(X*Y + bias), X'dims >=2 and Y'dims shoule be 2.
2. Act currently only be supported ReLU. (Will add GeLU in the future).

* Added UT to fused_gemm_epilogue op.

* Added LinearAct Pattern

1. Added LinearAct into graph_pattern_detector.* to define (2.)'s
pattern.
2. LinearAct is used to detect act(element_add(matmul_v2(x, w), bias)).
3. act currently only support ReLU (Will support GeLU in the future).

* Added FuseGemmEpiloguePass

1, Added FuseGemmEpiloguePass to handle nn.Linear + Act{ReLU}
fusion (GeLU will be supported in the future).
2. Only support matmul_v2 from nn.Linear.

* Added pybind to BuildStrageter.fuse_gemm_epilogue_.

* Added UT for fuse_gemm_epilogue_pass.

* GeLU support and EpilogueSingleton

1. Added GeLU support to fused_gemm_epilogue op.
2. Added EpilogueSingleton to cache auxiliary pointer.
3. Added related UTs.

* Rename cublaslt_epilogue_opto gemm_epilogue_op.*.

* Added both train and infer pattern to LinearAct.

1. Added support of fwd graph with grap_ops linking to LinearAct.
2. Added related changes to fuse_gemm_epilogue_pass for above
modification.

* Changed CUDA requirement from 11.4 to 11.6 for fuse_gemm_epilogue_pass.

* Added identity activation support to gemm_epilogue_op.

* Added Linear Fusion (matmul_v2 + ele_add)

1. Added matmul_v2 + ele_add pattern to LinearActPattern.
2. Added matmul_v2 + ele_add support to fuse_gemm_epilogue_pass.

* Rename gemm_epilogue_op.* to fused_gemm_epilogue_op.*

* Add fused_gemm_epilogue_grad op.

1. Added fused_gemm_epilogue_grad to support backward epilogue fusion.

* Add UTs to fused_gemm_epilogue_grad_op.

* Change attribute name in fused_gemm_epilogue_grad_op for clearing.

* Allow DX and DBias be dispensable to fused_gemm_epilogue_grad op.

* Added ElementwiseAdd+Matmul+Act graph pattern detection.

* Fuse backward of Linear( Act(x))

1. Added backward fusion pass to Linear( Act(x)).
2. Added backward fusion pass to Linear(x).

* Added UTs to backward fusion of Linear(Act(x)).

* Complete document of arguments to fused_gemm_epilogue_op.

* Made arguments of some functions pass by reference.

* Modify code with review comments.

1. Made arguments of some function pass by reference.
2. Removed redundant code.
3. Followed Google code style to change code.

* Made 'const' code style be consistent

* Fixed random seed of python UTs.

* Set Compiling constrains to cuBlasLt

1. Require CUDA 11.6+
2. Remove fuse_gemm_epilogue related tests when CUDA < 11.6.

* Code Reivew from Paddle

1. Changed arguments name is_first_gemm to without_x_gradient for
clearing.
2. Applied PADDLE_THROW in fused_gemm_epilogue_op.

* Remove EpilogueSingleton

1. Applied ReserveSpace to replace Epilogue for passing auxiliary
pointers between FWD and BWD.

* Fix a logical error and enhance UTs.

1. Added act op count checking in UTs.
2. Fix issue to fuse backward or ReLU(Linear(X)).
3. TODO: solve GELU fusion issues.

* Fix Linear and GeLU fusion issues.

1. Modified graph_detech_pattern to fit with both linear wiht gelu or
relu.
2. Modified data range in Uts to allow negative values.

* Removed fused_gemm_epilogue_op.h.

* Rename namespace pten to phi.

* Rename name of arguments in fused_gemm_epilogue_op

1. bias -> Bias.
2. out -> Out.
3. reserve_space -> ReserveSpace.

* Change EpiloguePassActivationCache as local variable.

1. Removed singleton in EpiloguePassActivationCache.
2. Made EpiloguePassActivationCache as an argument to each pass
functions.

* update cmake file to remove fluid kernel

* add pten declaration.h to where pybind.h used

* fix sync_bn and tensorrt_engine

* refine detection_library

* fix interpreter_core

* support eager legacy

* fit eager legacy for pten

* fall back to cpu if not found kernel

* fix compile problem

* fix compile problem

* refine fallback logic

* fit operator.run()

* fix xpu compile

* fit for new_exec

* add REGISTER_OP_WITHOUT_GRADIENT

* un-cache pt_kernel_context

* fix compile

* fix cudnn

* fix compiling with on_infer

* fix mkldnn

* fix isfinite_v2

* fix xpu problem

* fix op_device

* refine fallback for xpu

* fix xpu compile

* merge develop

* refine code format

* fix compile

* fix compile

* add data_transfer

* fix PreparePtenData

* fix cpu context

* merge develop

* fix compile

* fix error device context

* fix xpu

* fix dev_ctx

将 #35905 和 #35843 PR中新增的的python api接口移到incubate目录下。

这个PR只包含fused_feedforward前向的代码。

相关kernel实现：fused_dropout_act_bias, fused_residual_dropout_bias, fused_layernorm_residual_dropout_bias

fused_feedforward是一个融合算子，该算子对transformer模型的feed forward层的算子进行融合和封装，使得前端只呈现一个接口，通过融合减少部分访存和kernel launch的时间，以此提升性能。

功能：本PR的目标是提高attention模块的计算性能。
为了减少框架层对op的调度开销，本PR通过在C++层手动实现attention模块，对外提供attention 大op；
为了减少防存开销，本PR采取了两种优化方法：
（1）在q,k,v计算时通过共享输入X，将该处的gemm，transpose和bias add从三次调用减少为一次；
（2）使用kernel融合优化技术，在不同cuda kernel之间通过寄存器传输数据；

Fused elementwise_add, dropout, elementwise_add and layer_norm into one operator, only support Forward. 
No Python API changed.

This PR adds fused transformer related files defining c interface including class, function etc..

* fix test_unpool_op

* fix test_inplace_addto_strategy

* fix test_conv2d_fusion_op

* fix test_imperative_lod_tensor_to_selected_rows, test_imperative_selected_rows_to_lod_tensor

* fix test_dot_op

* fix test_correlation_op

* fix tracer

* fix test_memcpy_op

* [ROCM] update fluid operators for rocm (part3), test=develop

* fix clang format error, test=develop

* added external reorder to profiler

* resolved conflict

* added enable_static

* initial version of lstm, not working yet

* added lstm to operators.cmake

* added vanilla lstm mkldnn op

* added peephole weights integration

* minor changes

* added formatting

* added fusion_lstm_mkldnn to static_whitelist

* added formatting

* removed comment

* moved use_peepholes attribute inside is_cached block

* reverted wrong changes

* minor formatting change

* minor changes

* changed stream handling

* minor change

* added datatype to GetExpectedKernelType()

* added reading stream from TLS

This reverts commit d834f4e6.

* added external reorder to profiler

* resolved conflict

* added enable_static

* initial version of lstm, not working yet

* added lstm to operators.cmake

* added vanilla lstm mkldnn op

* added peephole weights integration

* minor changes

* added formatting

* added fusion_lstm_mkldnn to static_whitelist

* added formatting

* removed comment

* moved use_peepholes attribute inside is_cached block

* reverted wrong changes

* minor formatting change

* minor changes

* Compiling operator libraries with Unity Build on Windows CPU.

* Compiling operator libraries with Unity Build on Windows GPU, no_test, test=windows_ci

* Add option in windows ci script, no_test, test=windows_ci

* Optimize parallel compiling, test=develop

* remove limit of parallel compile and skip some ops in UB, test=develop

* remove changes of header file, test=develop

* remove changes of header file, test=develop

* fix test_eye_op unittest failed, test=develop

* Compiling operator libraries with Unity Build on Linux, test=develop

* set default WITH_UNITY_BUILD=OFF, test=develop

* Move unity build rules into a single file and add comment, test=develop

* optimize parallel compilation, test=develop

* fix undefined reference error on coverage ci, test=develop

* skip_layernorm_op done

* add unittest

* slice op convertor support trt < 6

* skip_layernorm only work in ernie

* add fused_bn_add_relu op

* Add oneDNN fusion_gru kernel and fix fc+gru pass
test=develop

* Formatting changes
test=develop

* Lint fixes
test=develop

* Add memory::format_tag::any to GRU weights
test=develop

* Fix build with CUDA

* Fix build with CUDA v2

* 1. add embedding eltwise layernorm fuse
2. add embedding eltwise layernorm op
3. refine inplace_add_relu
4. refine fc_eltwise_layernorm
test=develop

* 1. refine fc
test=develop

* fix comments
test=develop

* fix comments

test=develop

* add bn and relu fuse pass

* add op attr assert and dtype assert

* fix some inputs&&outputs bugs for the fused op and pattern.

* add the unittest for fuse_bn_act_pass. test=develop

* use normative enforce statements. test=develop

* add the cpu test. test=develop

* add the support of batch_size=1 for the bn with relu op. test=develop

* add the error type for paddle throws. test=develop

* add fused_batch_norm_act and fused_batch_norm_act_grad to op_has_unsed_vars_white_list. test=develop

* Add the dynamic load of nvrtc, and support runtime compiling of CUDA kernel using nvrtc.
test=develop

* Call CUDA driver api to launch the kernel compiled by nvrtc.
test=develop

* Disable for mac and windows.
test=develop

* Refine the codes to support manually specified num_threads and workload_per_thread.
test=develop

* Refine the CUDA kernel to support large dims.
test=develop

* Add DeviceCodePool to manage all device codes.

* Add the first implementation fusion_group op.

* Add unit-test for fusion_group op.

* Add the check of result.

* Add the check of nvrtc in unit-test.
test=develop

* Add comment to explain the inputs, outputs and features of fusion_group op.
test=develop

* Disable fusion_group op for mac and windows.
test=develop

* Make the compiling of device code return status instead of hanging up.
test=develop

* Add the check of whether there is CUDA driver library, and do not core dump when failing to call the CUDA driver API.

* Unify fusion_group_op's input and output names.
test=develop

* Add the check of CUDA driver library in unittest.
test=develop

* Refine the calling of PADDLE_ENFORCE.
test=develop

* Move the codes of fused operators to operators/fused directory.
test=develop

* Correct the op name in cmake.

* Change the use of PADDLE_ENFORCE.
test=develop

* Add fc_elementwise_layernorm_fuse pass and unittest.

* Add fused_fc_elementwise_layernorm op and its GPU kernel.
test=develop

* Apply fc_elementwise_layernorm_fuse_pass to GPU inference.

* Add the setting of attrs in the definition of binary_op.
test=develop

* Add comment.

* Implement the unittest.
test=develop

* Change the unittest name of layer_norm.
test=develop

test=develop

* Inception fusion operator.
* Support horizontal layer fusion in conv_fusion_op.
* Search conv algo strategy for variable-length input.
   search N times and cache the searched algos. For other input, choose the algo of input whose area is closest to this input.

* Transpose-Flatten-Concat fusion operator.
* Add unit testing and fix bug.

* wip simplify operator framework

* wip

* wip

* done test=develop

* clean test=develop

* fix test=develop

* fix deps test=develop

* fix cpu build test=develop

* fix tensorrt build test=develop

* fix tests test=develop

* fix test=develop

* fix cpu build test=develop