* notest;test=model_benchmark_ci

* notest;test=model_benchmark_ci

* fix

* add elementwise max grad op for npu

* add elementwise max grad op for npu

* add elementwise max grad op for npu

* add elementwise max grad op for npu

* add elementwise max grad op for npu

* polish the error message of paddle.slice.

* polish code.

This reverts commit ac33c0ca.

[oneDNN] disable caching oneDNN primitives in  matmul v2, Reduce grad and elementwise_add grad, expand_v2 (#35132)

* - grad caching disabled of matmul_v1

- compilation fix

- compilation fix

* - reduction removed

* - Matmul v2 disabled caching

* Draft of further changes

* - workaround for reducegrad

* - fixes to UT

* - fix to compilation

* - another fix

* - fix

* add dlpack api and fix a from_dlpack 

* fix assign bug support fp16 uint8

* fix dygragh assign bool bug

* modify code style

* revoke bool modification

* gc for newexecutor, test=develop

* refine, test=develop

* add interpretercore_gc_helper.h,test=develop

* backup

* gc whit thread and device_event, test=develop

* refine, test=develop

* refine, test=develop

* refine, test=develop

* refine, test=develop

* fix bug, test=develop

* refine, test=develop

* refine, test=develop

* refine, test=develop

* add CheckGC, test=develop

* Support dropout backward in eval mode

* add downscale case

* minor fix

* minor fix

* polish code

* polish code.

* polish code.

* polish code.

* polish code.

* Modify into QueueSync QueueAsync

* fix complie on MacOS

* fix pointer

* fix conflict

* polish unittest

* fix windows fetch error

* polish code according reviewer

* fix device_guard on CPU place

Describe

Add feed_forward for fused attention op.
(1) Encapsulate matmul impl (forward and backward) used in attention op.
(2) Implement bias_add (forward and backward) used in attention op.