[Fix] Fix missing post_micro_batch_forward invocation in vl model

[nil0x9]

Currently vl base model does not override post_micro_batch_foward method, meaning that they inherit the behavior from BaseModel directly. When one trains a vl model with MoE text part, this would cause:

1. missing maxvio from tracking logs;
2. (critical) if NoAuxRouter is used, the bias would never be updated.

This PR fixes this issue.

[nil0x9]

@claude review

[claude]

Claude: ## Summary

Clean, correct fix. The BaseComposeModel was inheriting post_micro_batch_forward from BaseModel directly, which skips the MoE-specific logic (maxvio tracking, NoAuxRouter bias updates) when the language model is an MoE model.

The fix delegates to self.language_model.post_micro_batch_forward(), consistent with the existing delegation pattern used by scale_and_reduce_grad() and build_loss_ctx_batch().

Verdict

APPROVE