Add is_model_splitted interface

[zhaixuejun1993]

This pull request adds support for detecting and handling split-model computation graphs ("splitted" models) in the OpenVINO GGML decoder integration. The changes introduce a heuristic function to determine if a computation graph is a split subgraph and propagate this information through the decoder's API and logic. This enables more robust and accurate model handling, especially for cases where models are partitioned or composed of multiple subgraphs.

Key changes include:

Split-model detection and handling:

• Introduced the is_model_splitted function in utils.cpp and its declaration in utils.h, which heuristically checks if a ggml_cgraph is a split-model fragment. This function is now used in the dynamic compute path to determine whether to use naive computation or not. [1] [2] [3]

• Updated the GgmlOvDecoder class and its constructor to accept and store a model_is_splitted flag, and added the is_splited_model() virtual method to the decoder interface. This allows downstream components to query if the current model is split. [1] [2] [3] [4] [5]

Integration with compute logic:

• Modified the dynamic and static compute routines in utils.cpp to utilize the new split-model detection logic and to correctly construct GgmlOvDecoder instances with the appropriate model_is_splitted flag. [1] [2] [3]

[wine99]

@zhaixuejun1993 The split check seems to be expensive when running with quantized models

llama-simple -m llama-3.2-1B-Q4_0: