OPT & Clean code of openvino_vision_embeddings_merger_model inputs processing

[zhaixuejun1993]

1. reconstruct function get_window_index(), avoid unnecessary memory copy & memcopy() replace loop copy;
2. removed unused variable & merge loop to improve perf;

[Copilot]

Pull Request Overview

This PR optimizes the OpenVINO vision embeddings merger model by improving input processing performance and cleaning up the code. The changes focus on eliminating unnecessary memory operations and improving computational efficiency.

Key changes:

• Replaced manual tensor copying with direct tensor construction using existing data
• Consolidated nested loops to reduce memory allocations and improve cache locality
• Removed unused variables and optimized data type usage

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
src/cpp/src/visual_language/qwen2vl/classes.cpp	Optimized tensor creation and merged position ID generation loops
src/cpp/src/visual_language/qwen2_5_vl/classes.cpp	Restructured window index function to eliminate unnecessary memory operations

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}