Add preprocessing utils for Qwen3-Omni

[hengtaoguo]

Description

• Add image/video/audio preprocessing utils for Qwen3-Omni in MaxText.multimodal.qwen3_omni_preprocessor.preprocess_mm_data_qwen3_omni(), returning dataclass Qwen3OmniPreprocessorOutput containing all preprocessed data (pixel_values, pixel_grid_thw, video_values, video_grid_thw, video_second_per_grid, audio_values, audio_mask).
• Add unit test comparing MaxText implementation with Qwen3-Omni's processor on HuggingFace.
• [WIP] Refactor [multimodal_utils.py]:
  • MaxText.multimodal.utils: Commonly used basic functions such as image loading and normalization.
  • MaxText.multimodal.{MODEL}_preprocessor.py: Model-specific preprocessing utils.
  • MaxText.multimodal.preprocessor.py: Centralized function preprocess_mm_data() will route to model-specific preprocessing logics based on model name.

Tests

Passing unit tests for MaxText preprocess_mm_data_qwen3_omni vs HuggingFace Qwen3OmniMoeProcessor:

python -m unittest tests.check_qwen3_embedding_vs_reference.TextQwen3OmniPreprocessing

Checklist

Before submitting this PR, please make sure (put X in square brackets):

• [x] I have performed a self-review of my code. For an optional AI review, add the gemini-review label.
• [x] I have necessary comments in my code, particularly in hard-to-understand areas.
• [x] I have run end-to-end tests tests and provided workload links above if applicable.
• [x] I have made or will make corresponding changes to the doc if needed, including adding new documentation pages to the relevant Table of Contents (toctree directive) as explained in our documentation.

[eitanporat]

is the functionality implemented on cpu in numpy in the torch variant. if so, is there a reason not to want to reuse it?

[eitanporat]

could you add the new requirements to the pyproject toml (decord and librosa)?

[github-actions[bot]]

🤖 Hi @hengtaoguo, I've received your request, and I'm working on it now! You can track my progress in the logs for more details.

[hengtaoguo]

is the functionality implemented on cpu in numpy in the torch variant. if so, is there a reason not to want to reuse it?

This has been a long-standing constraint, we intentionally exclude torch from our dependency. So we cannot use torch resize functions and need to reimplement everything in numpy/jnp.