[bugfix] fix gather_from_sp

[Jintao-Huang]

No description provided.

[gemini-code-assist]

Code Review

This pull request refines sequence parallel operations across several model files. Key updates include replacing reduce_scatter_to_sequence_parallel_region with scatter_to_sequence_parallel_region, which removes the need for manual scaling by the tensor model parallel size. Additionally, the gather_from_sequence_parallel_region calls now explicitly set tensor_parallel_output_grad=False, and forward passes in Qwen models are wrapped with a CUDA RNG tracker fork to ensure consistent randomness. I have no feedback to provide as there were no review comments to evaluate.