fix(trainer): support selective logits with sequence parallel

[zjn20030811]

Summary

Fixes #9765 by enabling use_logits_to_keep with sequence parallel SFT.

• derive a local boolean selection after SP causal label shifting;
• run the language-model head only for supervised (or shared suffix) positions;
• scatter per-token losses back into the full local frame before GatherLoss, preserving SP/ring ordering and normalization;
• keep loss scaling, packed boundaries, token accuracy, and all-ignored shards aligned;
• retain the existing full-logit path for custom loss functions, label smoothing, and channel loss until those paths expose SP-aware contracts;
• keep RLHF trainers on their established full-logit path through an explicit capability guard.

Validation

• py -3.12 -m pytest -q tests/test_align/test_sp_logits_to_keep.py tests/test_align/test_rlhf_loss.py
• py -3.12 -m pre_commit run --files swift/trainers/mixin.py swift/trainers/seq2seq_trainer.py swift/trainers/utils.py tests/test_align/test_sp_logits_to_keep.py
• py -3.12 -m compileall -q swift/trainers tests/test_align/test_sp_logits_to_keep.py

The focused regression suite covers single-row masks, batched suffix selection, loss/gradient equivalence, all-ignored local shards, packed sequence boundaries, accuracy reconstruction, and the non-SFT capability guard.

[zjn20030811]

Validation update: the focused sequence-parallel suite covers masked rows, packed boundaries, gradient/loss equivalence, all-ignored shards, accuracy reconstruction, and the non-SFT capability guard. The branch is ready for review.

[tastelikefeet]

When will this code be used?