(feat): agentic rl stepwise learning GiGPO, support for multi-modal distill

[PanAndy]

(fix): fix agg loss bugs.
(docs): Add documentation for the distillation pipeline and update da…
(fix): fix seed in data generator.
(chore): update dpo_config.yaml file.
(feat): agentic add gigpo.
(docs): add documentation for supporting new models.
(feat): add support for multi-modal distill.

[CLAassistant]

All committers have signed the CLA.

[PanAndy]

In this PR, we added StepEnvManager and implemented the GiGPO algorithm.

We removed the explicit addition of and tokens in the response. This was because we found that adding could introduce abnormal logprobs values, and this extra text should not be included in the loss calculation.
Upon verification, even after removing the explicit addition of , the model was still able to learn the desired format.

The training curves are shown below:

NaN appeared in the later stages of stepwise training.
We are very curious why tokens with such small probabilities are being sampled, and will continue to investigate this issue.

[ZhangXcho]

能否提供gigpo的训练入口呢

[HuangJoJo]

能否提供gigpo的训练入口呢

可从qwen2.5-0.5B-agentic_sokoban示例入手