feat: support Qwen3.5-VL model on npu device[6/N].

[yingxudeng]

No description provided.

[gemini-code-assist]

Code Review

This pull request adds support for the Qwen3.5 VL model, including updates for linear attention, multimodal Rotary Positional Embeddings (mRoPE), and deepstack processing. It also refactors model type detection and enhances NPU attention layers. Review feedback highlights a critical compilation error from modifying a constant reference, a memory allocation mismatch in the KV cache, and a style guide violation regarding anonymous namespaces. Additionally, suggestions were made to fix hardcoded logic in RoPE and deepstack retrieval to ensure safety and configuration compliance.

[wly-115]

torch::Tensor deepstack_process(torch::Tensor hidden_states, torch::Tensor visual_pos_masks, const torch::Tensor& visual_embeds)
I remember Qwen 3.5 already removed DeepStack.