[Fix] Forward packed dataset mode to GPT-OSS preprocessors

[primorLee]

Summary

• forward use_packed_ds from LazySupervisedDataset to the selected preprocessor
• cover single-image, multi-image, video, and pure-text supervised samples
• add a regression test that checks every modality forwards the dataset's packed mode

Problem

For Qwen3-style SFT, get_preprocess_function() selects preprocess_internvl2_5. That preprocessor only disables per-sample max-length padding when group_by_length or use_packed_ds is true. Packed training intentionally does not rely on group_by_length, but the dataset did not forward self.use_packed_ds, so each item could be padded to tokenizer.model_max_length before reaching PackedDataset. This defeats variable-length packing and changes the effective token composition.

Validation

• python -m pytest tests/test_gpt_oss_packed_preprocessing.py -q (1 test, 4 subtests passed)
• python -m compileall -q ... on Windows Python 3.12
• the same regression test and compile check in a clean Linux Python 3.10 container
• the new test fails against the current main source because all four call sites omit the flag
• Ruff passes for the new test; the modified legacy script retains the same seven pre-existing findings as main

Closes #1166