修复了fused_transformer_encoder_layer fine-tune过程发现的一些问题：

    fused_attention_op添加attn_mask=None的支持：PR
    pre_layer_norm处理问题：PR
    参数处理，计算错误的问题：PR
    add_bias计算错误问题：PR
    添加pure fp16的支持：PR

* Revert "Align CTC grad scale same with ESPNet (#34729)"

This reverts commit 10f9644c.

* ctc grad compute on gpu

* fix BatchNorm for fp16

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；

* add op: fused_feedforward(backward) (#35611)

这个PR是fused_feedforward反向的代码

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。

* Move fused_attention and fused_feedforward functional api path to incubate (#36704)

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

This is a fusion operator to compute feed forward layer in transformer model architecture.

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；

Add paddle.nn.functional.sparse_attention API

    本个PR主要将sparse_attention功能在python层进行了一层封装，OP的主体代码见：#PR35676

    此外，对于封装的python 接口，增加了相应的单测。

* fix pad tuple

* fix format

* fix unpool doc, test=document_fix

* fix typo for python example, test=document_fix

* Change the invoking method of settiem from numpy to set_value op when value is not tensor

* fix the check logic for inplace in setitem

* fix the unittest problem caused by setitem doesn't support fp16

* modify some code format in setitem

* [NPU] fix depthwise_conv2d_grad, test=develop

* remove debug files, test=develop

* Add New Op: gumbel_softmax

* Add New Op: gumbel_softmax

* Add New Op: gumbel_softmax (amend)

* add __main__ function in unit test

* fix bugs when test in windows ci

* update en docs

* delete reletive error in unit test

* delete relative error in unit test

* set hard=True in unit test

delete main function

1. optimize the error message of softmax_with_cross_entropy_op;2. add input value check for cross_entropy, if the dimention of input is zero, raise error. test = develop

* fix interpolate launch error, test=develop

* fix area mode for interp, test=develop

* [RC22] Fix linear with matmul_op replace

* [RC22] Fix linear with matmul_op replace

* [RC22] Fix linear with matmul_op replace

* [RC22] Fix linear with matmul_op replace

* [RC22] Fix linear with matmul_op replace

* add conv op check for illegal input or attributes

* add kernel, stride check

* add unitest for param out of range

* delete max limit check

This PR adds error exception in fused transformer python interface.
The function body are not implemented (will be implemented later).
Following zhiqiu's comment in previous PR-35206 (merged already), it is better to raise an exception instead of using "pass".

This PR adds fused transformer python related files. It defines interface of fused transformer.

Fused transformer implements an optimized version of transformer layer (in python/paddle/nn/layer/transformer.py). In this PR, four layers (functions) are defined:
(1) FusedMultiHeadAttention: multi-head attention layer
(2) FusedFeedForward: feed forward layer
(3) FusedTransformerEncoderLayer: transformer encoder layer
(4) FusedTransformer: transformer layer

* add maxunppol2d op, test=develop

* fix typo, test=develop

* fix unpool unitest, test=develop

* fix unpool code-example, test=develop

* fix for unpool_op_unittest,test=develop

* fix example code, test=develop

* add noqa:F401, test=develop

* fix converage, test=develop

* fix unitest for unpool, test=develop

* rename unpool2d to unpool, test=develop

* rename unpool2d to unpool, test=develop