Fix attention size computation error in OpenVINO backend for LLM

[zhaixuejun1993]

This pull request introduces several updates to the OpenVINO GGML decoder logic, primarily improving the detection and handling of attention-related operations and key-value cache identification. The changes enhance robustness when parsing computational graphs and ensure that certain preprocessing steps are only applied in appropriate contexts.

Improvements to attention and KV cache handling:

• Improved the logic in compute_llm_params to check for deeper source validity when handling GGML_OP_SOFT_MAX, preventing potential null pointer dereferences.
• Added new logic to extract attention_size from specific GGML_OP_MUL_MAT graph patterns involving permute and view operations, increasing accuracy in parameter computation for attention mechanisms.
• Refactored the is_kvcache static method to prioritize buffer usage checks, making the order of conditions more robust and consistent.

Preprocessing logic improvements:

• Updated the preprocess function to only add sliced masks when the decoder is stateful, preventing unnecessary operations for stateless models.## Overview

Additional information

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: