[cherry pick] Support optional residual add in fused ops and slice large tensor for cudnn_softmax (#43719)

 [cherry pick] Support optional residual add in fused ops and slice large tensor for cudnn_softmax

cherry-pick #43635 #43681 #43474

* Rename dropout is test (#43098)

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

* fused_attention fused_feedforward api support Model Tensor Parallel (#42985)

* fix is_test bug in fused_feedforward. (#43508)
Co-authored-by: NLi Min <11663212+limin2021@users.noreply.github.com>

* #1 migrate dist-related type()-> dtype()

* move datatype function from pten -> fluid/framework

* change type() in imperative into convert(dtype())

* modify xx_tensor->type into xx_tensor->dtype

* change the set_type interface and the caller

* modify xx_tensor.type into xx_tensor.dtype

* fix mutable_data(place, dtype())

* change caller of mutable_data in pten and distributed

* change the caller of mutable_data in fluid/framework

* change the caller of mutable_data in imperative directory

* mutable_data: inference

* update the call of mutable_data

* transfer MakePenScalarArray MakePtenScalar ResetHolderWithType

* pass the compile. the next step is remove VarType in Pten

* fix all and remove VarType from pten. success in linux. Next task is other platform

* fix conflict with develop

* fix compiled error

* Fix reset conversion

* fix conflict

* fix compiled problem

* fix typo

* Fix << in tensor_utils.cc

* fix type->dtype

* fix unittest

* fix tensor init constructor

* fix DataTypeSize for BFloat16

* fix code style

* fix npu compiled error

* fix npu

* compile npu sucessfully

* fix conflict

* fix conflict
Co-authored-by: Nxiongkun <xiongkun03@baidu.com>

Add support for bias is none for fused_attention op.

att, bug fix

目前的fused_attention_op不支持attn_mask=None的输入，本PR对此进行了补充，并补充了相应的单测逻辑。

* Fix bug when pre_layer_norm is false.

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；

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；