* support bfloat16 for amp_decorate

* add check_finite for bf16

* fix bug

* add ut

* add ut

* refine code

* fix arm fp16 compile error

* polish macro impl

* fix distributed_fused_lamb nan

* remove CUDA_ASSERT

* add reduce_mean,reduce_sum primitive ops

* add ne_p gt_p primitive operators

* add ge_p abs_p primitive oparators

* normalize yaml op label

* revert op_compat yaml change

* fix prelu and rnn compat problem

* replace api by op

* optimize executor log

* delete log in new exe

* add log for old executor

* use LOG_FIRST_N(INFO, 1)

* clear extra attr of activation op in opmaker

* fix syntax bug

* fix mkldnn kernel

* fix merge conflict

* fix bug

* add softmax infer kernel

* add scope cache & reuse

* add gc scope for end of each train step

* del scope reuse for jit

* refine code

* test

* make memcpy op to support custom device

* fix bug

paddle::platform::CudaAtomicAdd
https://github.com/PaddlePaddle/Paddle/issues/45881

* fix softmax int64

* follow comments

* fix split bug in static mode

* modify code style

* modify code style

* add unit test for split

* Allow manaully set py_reader name in standalone executor

* Fix CI errors

* add paddle.incubate.sparse.is_same_shape

* add paddle.incubate.sparse.is_same_shape

* add paddle.incubate.sparse.is_same_shape

* add paddle.incubate.sparse.is_same_shape

* fix doc

* add operator<< for BuildStrategy

* add fake_coalesce

* fit allreduce mode for new_exe

* remove dubeg code

* follow comments

* migrate load kernel

* remove load op

* fix test failed

* [CustomDevice] add dy2static support

* update