* Add bf16 uniform random initializer

* Remove duplicated section

* Change UT to CPU place only

* Put detail functions into anonymous namespace

This is cherry-pick of #32281

cherry-pick of #32666

后续修复计划是啥

- Executor is nt always having FLAGS_use_mkldnn set to true

* optimize update_loss_scaling_op by fused for loop to one kernel, test=develop

* remove useless while loop and optimize variable name, test=develop

* optimize variable name from out_addrs_tensor to out_addrs_mem, test=develop

* optimize variable name for readable by change prefix identifier from t_ to local_

* [OPs] Bug fix, fix the segment mean for illegal syncthreads usage.

This reverts commit 4b7242b0.

Co-authored-by: NYang Zhang <yangzhang@live.com>

* add sendrecv, test=develop

* clear CUDA compile environment on windows

* fix Windows CI

* fix Windows CI

* fix Windows CI

* new optimize for where_index_op with prefix sum version.

* write a scan prefix sum kernel with stream for where index op.

* optimize where_index by using cub::DeviceScan::InclusiveSum instead of imperfect self-kernel.

* remove CheckTrue struct and rename stide_array for readable.

* optimize variable name for readable.

* optimize function name and annotation.

* optimize pull sparse

* optimize pull sparse

* change macro

* format

* Unset ReserveSpace for inference program.

* Support training from an inference program.

* fix model parallel

* rm parallel_help.py

* add embedding

* support backward return None.

* edit unittest.

* edit code according to CI

* Improve error information

* optimize slice op and slice grad op, test=develop

* optimize variable name and annotation information, test=develop

* skip op has no fp16 kernel

* add ut

* fix airank bert emb order

* move input num check to converter

* add input num check

* add unused var check white list

* add trt runtime version check

* use different wrap, and change to major version check

* update

* fix bug: when x.dim < y.dim, the result of compare_op is inverse to expected result

* support the cuda for fix the compare broadcast bug