Fix Qwen3-VL Processor flattening multi-image batches (fix #41709)

[Elon7069]

Fixes #41709

This PR fixes how Qwen3-VL Processor handles multiple images per sample.
Previously, all images were flattened across the batch, causing shape mismatches in model forward passes.

Changes:

• Updated Qwen3VLProcessor to preserve per-sample image grouping.
• Added test to verify batch dimension consistency.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: qwen3_vl

[Elon7069]

Hi, this PR fixes the issue where Qwen3VLProcessor was flattening multi-image batches. It now preserves per-sample grouping and passes the added unit test.

[Rocketknight1]

cc @zucchini-nlp

[zucchini-nlp]

In Qwen models, the pixel shape must be NxD. We don't need a PR, thanks for working on it though