feat(sft): support logits selection with Ulysses sequence parallelism

[Excelius-Wang]

Sequence-parallel SFT currently rejects explicit use_logits_to_keep=true. Add opt-in selection after SP has shifted and sharded labels, retaining a sentinel on prompt-only ranks. Restore scalar token losses and integer predictions to their local positions, then reuse existing gathering, weighting, normalization and accuracy. Evaluation keeps full logits; shared DPO/KTO/GKD preparation is unchanged.

Support covers native tensor-selection text causal LMs and text-only Qwen3.5/3.6 MoE inputs, with Ulysses, per-device batch one and no packing/padding-free. Media inputs, other multimodal architectures, Ring, custom loss, label smoothing, Unsloth and Liger remain excluded. Accept the all-zero modality IDs emitted by the Qwen text collator while rejecting nonzero media IDs.

Related #9765. This implements the reporter's model-family text SP path, but does not establish that their full 35B/128K, eight-H20, ZeRO3 workload fits in memory.

Validation:

• 23 tests and 8 subtests passed. Real two-process CPU Gloo/Ulysses/DDP on tiny Qwen3 and Qwen3.5 MoE, including full training and LoRA/checkpointing, compares loss/all trainable gradients with independent full-sequence references across 72 combinations; also tests auxiliary router-loss parity, accuracy, channel loss, exclusions and shared/non-SP regressions.
• The same MoE test on clean main fails at its original NotImplementedError.
• Normal SwiftSft with Qwen3.5 MoE, real template/tokenizer, SP2, LoRA, checkpointing, weighted examples, accumulation two and per-step evaluation completes two steps. Selection on/off yields identical train/eval losses and all 91 final state tensors; adapters are updated.
• Full pre-commit run --all-files and staged diff checks pass.
• Separate single-H100 BF16 native model test with LoRA/checkpointing and SGD passes loss/gradient/update parity; this was not multi-GPU SP or trainer integration.

CPU integration uses torch 2.9.1 and transformers 5.12.1. No ZeRO/FSDP, full pretrained long-context, image/video or GPU performance claim; full upstream suite not run.