Enable minicpm3 for OpenVINO backend architecture tests

[zhaixuejun1993]

This pull request introduces improvements to how sequence and attention parameters are handled in the OpenVINO decoder implementation. The main focus is on ensuring that parameters such as token_len_per_seq and attention_size are correctly set and passed as inputs when available.

Parameter handling improvements:

• In compute_llm_params, the token_len_per_seq parameter is now set based on the size of node->src[1] if it hasn't been set yet and the source exists. This ensures accurate tracking of token lengths for each sequence.
• In add_extra_inputs, both attention_size and token_len_per_seq are now only added as extra inputs if their values are set (not equal to -1), preventing the creation of unnecessary or invalid inputs.## Overview

Additional information

Requirements

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