fix(loss): reuse aligned cross entropy under sequence parallelism

[Excelius-Wang]

Explicit --loss_type cross_entropy recomputes ordinary causal cross entropy from labels that sequence parallelism has already shifted and sharded. This shifts the targets again and loses the gathered loss; with token weights, the gathered weights can also have a different length from the locally recomputed loss.

Reuse the per-token loss already aligned, gathered and weighted by Seq2SeqTrainer. When the callback infers the token count, preserve the default trainer's distributed normalization and training/evaluation scaling. The non-sequence-parallel path is unchanged.

Validation:

• CPU Gloo regression coverage for Ulysses, Ring and hybrid layouts, plus combined data/sequence parallelism with unequal valid-token counts, aligned/tail-padded/packed sequences, absent/nonuniform/zero weights, explicit/inferred counts, token averaging on/off and training/evaluation. Compare losses and gradients against both the default trainer path and an independent PyTorch CE reference (288 combinations).
• OMP_NUM_THREADS=1 /tmp/ms-swift-core-test-venv/bin/python -m pytest tests/sequence_parallel/test_custom_cross_entropy.py tests/train/test_cross_entropy_loss.py -q: 5 passed, 8 subtests passed.
• /tmp/ms-swift-lint-venv/bin/pre-commit run --all-files: all hooks passed.
• Additional local probe using Swift Qwen3 text-template encoding and tiny randomly initialized Qwen3 logits reproduces the original loss mismatch and weighted shape error; the fix matches default loss and local logits gradients.

Coverage uses CPU Gloo with one or two data-parallel replicas. It does not exercise GPU distributed attention or a full SFT run. Related #10100 added token weighting for custom CE but did not validate sequence parallelism.