Merge pull request !3846 from liuxiao93/Add-ops-SeluSquaredDifference

Merge pull request !3547 from wangcong/branch_07/27

for bert_thor 2nd

for bert_thor 3rd

for bert_thor 4th

Merge pull request !3425 from fangzehua/avgpool

Merge pull request !3324 from fangzehua/reduce_any

Merge pull request !3247 from DeshiChen/0720_akg_registers

AddN/Equal/GreaterEqual/Less/LessEqual/Log/Neg/Pow/Reciprocal/ReduceMax/ReduceMin/ReduceSum/Square.

Merge pull request !3175 from fangzehua/scatter_add_vm

Merge pull request !3158 from DeshiChen/0717_regop_gelu

Merge pull request !264 from wuxuejian/padding_op

Merge pull request !3147 from yanzhenxiang2020/add_rnnt_cate_open.new

Merge pull request !3135 from VectorSL/cast2

Merge pull request !3050 from liuxiao93/Tan/TruncateDiv/TruncateMOd

Merge pull request !265 from yanghaoran/incubator

Since akg supports both Ascend and Gpu, but their supported type and format are different,
so we use two directory "ascend" and "gpu" to store their registers respectively, and use
an attribute "processor" to distinguish them.

Main changes:
1) Add two op register class "AkgAscendRegOp" and "AkgGpuRegOp", inherited from the original AkgRegOp.
2) Rewrite akg ascend op registers with new interface, move them into directory "ascend".
3) Rename the imply_type from "AutoDiff" to "AKG".
4) Modify function FindOp, check the processor when imply_type is "AKG".
5) Modify function CheckRepetition, remove the judgement for impl_path, check processor instead.

TODO: Remove op registers in akg root path.

Merge pull request !2884 from liuxiao93/UnsortedSegmentProd

Merge pull request !2839 from jiangjinsheng/vm_population_count