fix: add preprocessing script for convert SenseNova-SI dataset to qwen3-vl format

[KetoneOlefine]

Summary

This PR fixes dataset schema incompatibilities that prevent Qwen3-VL data loading for SenseNova-SI-800K.jsonl.

Root causes

1. image has mixed types across samples (list[str] and str), which breaks Arrow/HuggingFace dataset construction.
2. Samples use conversations with plain string payloads, while the training loader expects:
   • messages key
   • content as a structured list of typed objects.

What changed

• Added training/qwen3_vl/preprocess_sensenova_si_dataset.py to normalize source JSONL:
  • image: str -> [str]
  • conversations -> messages
  • role mapping: human -> user, gpt -> assistant
  • text mapping: value -> [{"type":"text","text": ...}]
• Updated docs:
  • README.md
  • README_CN.md
• Updated example dataset config:
  • training/qwen3_vl/data.yaml now points to SenseNova-SI-800K_qwen3vl_format.jsonl

Fixes #31