Add Initial Qwen-2.5VL Image Processor Support

[sayanshaw24]

Description

This PR introduces support for Qwen2.5-VL image preprocessing in ONNX Runtime Extensions. It implements the full resize (including smart resize) → rescale → normalize → patching pipeline required by Qwen2.5-VL-7B-Instruct.

Key updates:

• New Qwen2.5-VL–compatible PatchImage op

  • Converts RGB HWC to CHW channel ordering
  • Performs accurate temporal padding and 9-D reshape/transpose logic
  • Produces Python-aligned patch embeddings

• Smart Resize support in the Resize op

  • Implements Qwen2.5-VL’s shortest-edge and longest-edge constraints
  • Supports pixel-count–based resizing (min/max pixels)
  • Matches smart_resize behavior from the HF transformers

• Qwen2.5-VL Normalization handling

  • Fixes channel-order expectations across resize → normalize → patch
  • Ensures exact mean/std behavior equivalent to HF transformers
  • Includes C++ normalization adjustments when Qwen2.5-VL mode is enabled

• Qwen2.5-VL processor JSON

  • Includes Qwen-specific resize parameters
  • Correct mean/std normalization
  • Patch dimensions (patch_size, merge_size, temporal_patch_size)

• RGB Correctness Guarantees

  • Fixes the BGR→RGB mismatch observed in PatchImage
  • Ensures alignment with the Qwen2.5VL

This enables ORT Extensions processing support for Qwen2.5-VL vision & multimodal models across ORT Extensions, ready-to-use in ORT GenAI. Note that we currently only provide single-image and non-video support.

Validation

• C++ unit tests verifying pixel_values match reference output; MSE comparison: achieved ≤ 1e-3
• Step-by-step parity checks across:
  • Resize
  • Rescale
  • Normalize
  • Patch extraction
• Regression-tested with Phi-3-V, Phi-4, MLlama, Gemma, etc. — no changes in their outputs

[tianleiwu]

We shall avoid such vector copy.

I think we can allocate with padded space. Fill them with padded value. Then apply HWC -> CHW (need to compute destination offset with padded height / width).