[ml service] Support flexible tensors for compatibility with filters like llama.cpp

[songgot]

• Refactor tensor allocation logic to support flexible tensor
• Enable compatibility with filters such as llama.cpp by allowing dynamic tensor management

[anyj0527]

Please rebase the branch c/src/ml-api-inference-single.c is updated in the recent PR merge.

[myungjoo]

Need a smaller unit test case. We execute all unit test cases for every PR, SR, build, ,..

[songgot]

Please rebase the branch c/src/ml-api-inference-single.c is updated in the recent PR merge.

Thank you. Rebase is completed.
For extensibility, we used format instead of is_flexible.

[songgot]

Need a smaller unit test case. We execute all unit test cases for every PR, SR, build, ,..
scenarioConfigLlamacpp TC is included in unittest_capi_service_extension.c
Generating TCs based on logic changes seems difficult.
Thank you.

[anyj0527]

LGTM👍 However there is a concern related with syncing nnstreamer to the tizen-unified. You may proceed the SR process of reflecting latest nnstreamer into the Tizen.

[jaeyun-jung]

well, should we keep current format here?
How about updating information in single instance as flexible?

// updating below information as flexible
GstTensorsInfo in_info;
GstTensorsInfo out_info; 

[songgot]

Currently updating within the sub-plugin. We need to fix all sub-plugins. Is it a necessary discussion?

executorch_llama_subplugin::getModelInfo (
    model_info_ops ops, GstTensorsInfo &in_info, GstTensorsInfo &out_info)
{
  if (ops == GET_IN_OUT_INFO) {
    in_info.num_tensors = 1;
    in_info.format = _NNS_TENSOR_FORMAT_FLEXIBLE;

    out_info.num_tensors = 1;
    out_info.format = _NNS_TENSOR_FORMAT_FLEXIBLE;

    return 0;
  }

  return -ENOENT;
}