[onert] Review internal custom operator support

[hseok-oh]

What

Review custom operator support and decide remaining or not

Why

We are supporting some custom operator internally to support converted tensorflow/pytorch model.
They may still need supporting, but may be converted to builtin OP by compiler now.

We didn't update custom operator support for a long time.
So if compiler support, it's better to update model by using latest compiler.

Previous issue: #15170

LIST

• AddV2 [onert] Remove AddV2 custom OP #15252
• ReduceAll [onert] Remove ReduceAll custom OP #15262
• MatrixBandPart [onert] Remove MatrixBandPart support #15317
• BatchMatMul [onert] Remove BatchMatMul custom OP #15314
• FusedBatchNorm [onert] Review internal custom operator support #15216 (comment)
• StatelessRandomUniform [onert] Review internal custom operator support #15216 (comment)
• Erf [onert] Review internal custom operator support #15216 (comment)
• DetectionPostProcess [onert] Review internal custom operator support #15216 (comment)

[hseok-oh]

AddV2

TFLiteConverter converts AddV2 to Add operation

https://github.com/tensorflow/tensorflow/blob/cb64295ec7308f770b22db6047a1e755b35b7bee/tensorflow/lite/toco/import_tensorflow.cc#L2574

ConverterMapType GetTensorFlowNodeConverterMap() {
  return std::unordered_map<std::string, ConverterType>({
      {"Abs", ConvertSimpleOperator<AbsOperator, kAnyNumInputs, 1>},
      {"Add", ConvertSimpleOperator<AddOperator, 2, 1>},
      {"AddV2", ConvertSimpleOperator<AddOperator, 2, 1>},

We can remove this custom op support.

[hseok-oh]

ReduceAll

TFLite support REDUCE_ALL builtin OP since v2.6

tensorflow/tensorflow@cb67fef

Converter: https://github.com/tensorflow/tensorflow/blob/cb64295ec7308f770b22db6047a1e755b35b7bee/tensorflow/lite/toco/import_tensorflow.cc#L2576

We can remove this custom op support

[hseok-oh]

Erf

Still converted to custom OP (v2.19)

[hseok-oh]

MatrixBandPart

Converted to builtin OPs

Example

We can remove this custom op support

[hseok-oh]

BatchMatMul

Converted to builtin OP
We can remove this custom op support

[hseok-oh]

StatelessRandomUniform

Still converted to custom OP

[hseok-oh]

DetectionPostProcess

Model: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/running_on_mobile_tensorflowlite.md

[hseok-oh]

FusedBatchNorm

Still converted to custom OP