Multi Resolution Support for Qwen2.5-VL Model

[quic-sanising]

Qwen2.5‑VL workloads often see variable image sizes; supporting multiple resolutions improves usability and benchmarking realism. This PR adds multi‑resolution support to the Qwen2.5‑VL specialization flow. The goal is to allow a single run to handle multiple image sizes more robustly while keeping specialization metadata consistent and avoiding shape/buffer mismatches.

What changed?

Qwen2.5‑VL specialization now supports multiple (width, height) pairs without requiring changes in model onnx:

• Accept width and height as either int or List[int].
• Reuse shared smart_resize utility instead of local implementation (imports: from qwen_vl_utils import smart_resize).
• Compute encoder specialization for each resolution by looping over (width, height) pairs.
• Compute decoder specialization based on the min_vision_size across all provided resolutions.
• Allow overriding image tokenization constraints via mm_processor_kwargs, particularly, min_pixels and max_pixels.
• Allow overriding prefill/decode vision_size specialization via vision_size user input.

Testing

• Single‑resolution path: pass height=int, width=int and verify behavior matches previous outputs.
• Multi‑resolution path: pass height=[...], width=[...] and confirm:
  • specializations are generated for each resolution,
  • min_vision_size is used consistently,
  • no shape mismatch at runtime.
• KV‑offload run with at least two distinct pixel_values shapes that are present in vision_session.allowed_shapes.

Checklist

• Support both scalar and list inputs for image resolution.
• Remove duplicated resizing logic in favor of shared utility.
• Select correct buffers for vision session based on input shapes.
• Pad vision embeddings to match language session constraints.