(sync): sync codes to github.

[PanAndy]

1. (feat): support pp + tie embedding.
2. (fix): Skip current step when final_response_mask.sum() == 0.
3. (fix): addressed all issues where the response mask sum was zero, both at the batch level and within individual domains.
4. (fix): fix mcore val.
5. (feat): add rlvr rollout pipeline for just val.
6. (fix): Fix VLM collator empty multi_modal_data.
7. (fix): remove fp8 quantization patch in vLLM 0.10.0.
8. (docs): Add docs for vlm rlvr pipeline.
9. (fix): set RAY_get_check_signal_interval_milliseconds to 1 to avoid queue exhaustion.
10. (fix): Fix the issue where DataProto could not collect metrics from a subprocess.
11. (fix): Fix math timeout and code extract.
12. (feat): support deepspeed ulysses.
13. (fix): fix vllm 0.10.0.
14. (docs): Add documentation for the DPO pipeline.
15. (feat): Optimize redundant imports and adjust timeout.
16. (fix): llm_as_judge with vllm strategy should init cuda before torch.random.manual_seed.
17. (docs): update readme taoSR1.
18. (fix): fix rlvr reward num_nodes infer.