support qwen3.6 grpo & in-place add lora

[hjh0119]

No description provided.

[gemini-code-assist]

Code Review

This pull request introduces a GRPO training script for the GSM8K dataset and implements support for Mixture-of-Experts (MoE) models within the Ray-based training framework. Key changes include specialized LoRA weight synchronization for MoE layers in the vLLM sampler, which now buckets expert weights separately and accumulates them on the CPU to prevent memory issues. Feedback was provided regarding a potential runtime error in the Megatron model logic where a variable could be accessed before being defined.