[Feature] Support balanced dataset to speed-up VL training

[yqyao]

Balanced Dynamic Mini-Batch for paper OmniBal: Towards Fast Instruct-tuning for Vision-Language Models via Omniverse Computation Balance.
Original Repo: https://github.com/ModelTC/OmniBal
speed-up results:
For 6 + 20b training, we can reduce training time from 14.5h to 9.5h with 64 GPUs using vit 9 llm 4096 input.

[hhaAndroid]

@yqyao Thank you very much for your open-source work. Could you please provide a sample configuration and a simple readme file?

[yqyao]

@hhaAndroid We have updated the config.