fix: harden multimodal dataset loading

[WangXukang-cypher]

Summary

Harden multimodal and omni dataset loading for calibration/test data.

This fixes MultiModalDataset when quant_algo is None, so sparsity-only or distill-only flows no longer crash on "int4_" in None. It also wraps non-user omni messages as text content instead of dropping assistant/system turns from the conversation.

Additional regression tests cover the quant_algo=None path, missing image error handling, non-list multimodal content, and preservation of assistant/system omni messages.

Test Plan

• python -m pre_commit run --all-files --show-diff-on-failure
  • Black: Passed
  • isort: Passed
  • Flake8: Passed
• python -m py_compile angelslim\data\multimodal_dataset.py angelslim\data\omni_dataset.py tests\test_multimodal_dataset.py tests\test_omni_dataset.py

[WangXukang-cypher]

Hi maintainers, just a gentle ping on this PR in case it got buried. No rush — would appreciate a review when you get a chance. Thanks!