make fused_multi_transformer support dynamically set the cache_kvs' shape and support input prefix_caches. (#46777)

* make fused_multi_transformer support dynamically set the cache_kvs' shape and support input prefix_caches.

* trim trailing whitespace

* fix `.cmake-format.py`

* revert npu ut changes, avoid npu ci error

* Add new attr of fused_multi_transformer

* fix format

* add note

* add in layer

* fixfixfixfix

* Support optional residual add in fused_attention and fused_feedforward.

* Add checkpoint and add the check of add_residual when pre_layer_norm is false.

* Add TODO and change the python api to add add_residual argument.

* use yapf to format all python file

* yapf exclude two unittests file for they rely on writing and reading file, and format will break them

* disable diff_py_file because too many diff files cause command following failed

* support weight transpose

* add ut

* add template

* fix transpose error

* fix transpose_comment

* add api tests

* add skipif

* add doc

* replace dropout_is_test with is_test.
* improve atol on a100.

* add fused_bias_dropout_residual_ln op and layer.

* refactor eager flags

* fix flags error when we switch from eager to dygraph

* fix ci problem

* fix ci

* fix ci

* merge develop and fix code style

* merge develop and fix code style

* fix op test error

* fix op test error

* fix op test error

* fix op test error

* fix op test error

* merge develop

* fix 'invalid escape sequence'

* fix assert error

* Fix bugs when bias is none for static graph for fused_attention op.

Add support for bias is none for fused_attention op.

fused_attention_op的实现中，使用了bias_add，且其实现是通过使用kernel primitive来实现的，之后kernel primitive的WriteData api接口及函数内部实现发生了更改，将判断越界的逻辑移到了template的参数中，使得调用的分支有错误，产生了越界赋值操作，污染了别的显存空间的内容。具体表现为：test_fused_attention_op_api.py 单次执行基本上不会报错，多次循环执行不同shape的输入，结果计算不对，具有偶发性，bug不易察觉。

* fix bug:
1. atten: set the default value of attn_dropout_rate to None
2. ffn: add activation parameter

* Fix fused_attention english doc test=document_fix

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

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；

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