Add ZeRO-3 leaf module support for Qwen MoE models

[HayatoHongo]

Summary

This PR adds explicit support for Qwen Mixture-of-Experts (MoE) models when running with DeepSpeed ZeRO-3.
When model.config.model_type == "qwen3_moe", the script sets the Qwen3MoeSparseMoeBlock as a ZeRO-3 leaf module.
This ensures that collective communication works correctly during training.

Changes

• Added lazy import of deepspeed and Qwen3MoeSparseMoeBlock inside the Qwen MoE branch.
• Applied deepspeed.utils.set_z3_leaf_modules() to register the MoE block.
• Added logging to indicate whether the setup was successful or skipped.

Notes

• This change does not affect non-Qwen models.
• If deepspeed or the Qwen MoE block import fails, the script logs a warning and continues without modification.
• No other functionality is changed.

Motivation

Without this fix, training Qwen MoE models under ZeRO-3 may hang or suffer from incorrect collective operations.
This patch enables stable fine-tuning of Qwen MoE models within the open-r1 training pipeline.

Limitations / Future Work

• This commit only covers Qwen MoE models; other types of MoE models are not yet supported.
• Incorrect collective operations may still occur for other MoE implementations.
• Future work should extend ZeRO-3 leaf module registration to other MoE model types.

Reference

A similar fix has been applied in the OpenRLHF repository:
OpenRLHF/OpenRLHF@d5fcb42#diff-da77c0ae1d958e6b8c491f9d6f1f8ad54ee9ab21c231d4b2490fb1c09af1046f