This PR creates a class to process the program at the C++ level. Currently, this class has one class method:
GetInputsOutputsInBlock()

* Add DCU backend support for custom ops

* Added checks for DeviceCopy and renamed some macros

[pass enhance] make the attribute check only object to which defined in op proto. test=develop (#34146)

* cache core.ops, test=develop

* refine, test=develop

* Refine Constructor logic of ParallelExecutor

* Replace executor into ParallelExecutor in run_program_op

* exclude lr scheduler's state from accumulators_holder

* fix when there is no learning rate scheduler

* make a copy of the loaded state dict to avoid modifying it

* add scatter-gather opt

* fix topo for pp

* rename function

* adam add input SkipUpdate

* add unittest

* add npu unittest

* fix xpu compile

* remove param stream

* kill process by group

* jit.save/load support method with parameters.

* add unittest and warning

* polish warning message.

As the title, add Python helper function to get block Inputs/Outputs. This function is helpful to distributed computing for sharding.

* Support Mac M1 make

* cmake version check

* Added OpTestTool for BF16 convenience

* fixes after review, names changed to snake case.

* fixes after review, naming reflects cpu.

* Support sccache to speed up compilation on Windows

* Support sccache to speed up compilation on Windows

* fix paddle.no_grad

* fix paddle.no_grad

* fix undef var

* fix undef var

* Fix gather_op by adding OurOfRangeCheck for param[Index]

* Code Optimization

* added printing tensor's format

* added suggested changes

* test=cpu-gpu-py3

* test=cpu-gpu-py3;notest

* notest;test=cpu-gpu-py3

* add npu operator for gaussian random.

* bugfix: add wait after memory copy.

* update gaussian random op: use TensorCopy.