[WebGPU] fixes pointer arithmetic bug for multi-image inputs for mistral3

[Prathik Rao (prathikr)]

Fix WebGPU multi-image Pixtral output handling

The customer's native multi-image repro shows that two-image Ministral/Pixtral inference succeeds on CPU and that each image succeeds independently on WebGPU, but the two-image WebGPU case fails in Generator.SetInputs, including repeatable 0xC0000005 native access violations.

PixtralVisionState ran vision once per image and wrote each result into an offset view of the combined image_features output. The WebGPU comment in src/ep/webgpu/interface.cpp identified the issue: p_device_ is a WGPUBuffer handle cast to uint8_t*, not byte-addressable memory; p_device_ + offset is therefore not a valid sub-buffer handle. The first image uses offset zero, while the second image does not.

This change runs each image into a standalone device output tensor and copies it into the combined feature tensor through ByteWrapTensor(...).subspan(...).CopyFrom(...). That path handles nonzero WebGPU offsets safely instead of constructing invalid handle-derived tensor views.

[Copilot]

Pull request overview

Fixes multi-image Pixtral/Ministral vision-output handling by avoiding invalid pointer arithmetic on non-byte-addressable device buffer handles (notably WebGPU), preventing Generator.SetInputs access violations when num_images_ > 1.

Changes:

• Stops constructing per-image output tensor views via feat_raw + offset, which is unsafe for EPs where p_device_ is an opaque handle (e.g., WebGPU).
• Runs the vision encoder into a standalone per-image output tensor and copies results into the combined image_features buffer using ByteWrapTensor(...).subspan(...).CopyFrom(...).

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.