[Feat] Add Qwen3.5 SFT example

[jayhenry]

Summary

• add an environment-driven Qwen3.5 SFT example for Alpaca and multimodal datasets
• default to 128K sequence length, SP=4, and soft packing without hard-coded storage paths
• add a single-node 8-GPU launcher and disable deterministic backward when FA3 is enabled
• support OpenAI-style video_url entries in Qwen3.5 multimodal tokenization
• mark empty text samples as damaged cache entries so JsonlDataset can filter them

Validation

• pre-commit hooks: passed (Ruff, formatting, Mypy, Pydantic checks)
• targeted tests: 3 passed
• config construction: verified both Alpaca and multimodal branches
• 8x H200 training:
  • 128K, SP=4 multimodal: completed one optimizer step without OOM
  • 16K multimodal: completed 20 steps without OOM or NaN/Inf
  • 16K Alpaca: completed 20 steps without OOM or NaN/Inf