fix(trainer): normalize partial gradient accumulation windows

[MrCapricornLiu]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

When the last accumulation window of an epoch has fewer micro-batches than
gradient_accumulation_steps, Trainer.compute_loss and
RerankerTrainer.compute_loss still divide their mean loss by the configured
window size. For models accepting loss kwargs, HF's training_step does not
divide that loss again. With a configured window of four and an actual window
of two, this produces half the full-batch reference gradient.

Use HF Trainer's current_gradient_accumulation_steps for that division, with
the configured value as a fallback for versions/callers without this attribute.
Full windows and the existing normalization branch conditions are unchanged.

The regression runs the actual HF training_step/Accelerate backward path
with Swift's Trainer and RerankerTrainer loss implementations, then compares
parameter gradients and accumulated loss with an independent concatenated-batch
reference. It uses a tiny CPU model, no model downloads. The test initializes
HF's training machinery directly and supplies Swift's required loss context;
it does not run the complete Swift initialization or epoch loop.

Experiment results

CUDA_VISIBLE_DEVICES='' PYTHONPATH=. python -m pytest -q \
  tests/train/test_trainer_partial_accumulation.py

• Original code: 4 failed, 5 passed. Failures are partial windows of one/two
  micro-batches in Trainer and RerankerTrainer.
• Fixed code: 9 passed with Transformers 5.3.0 / Accelerate 1.14.0;
  9 passed with Transformers 4.57.1 / Accelerate 1.11.0.
• Controls cover full windows and Trainer's model-without-loss-kwargs path.
• Changed-file pre-commit, flake8/isort/YAPF and git diff --check pass.

These are CPU backward tests with equal-size micro-batches, not distributed,
AMP, full fine-tuning or variable-batch weighting tests. The new pytest module
must be invoked explicitly; the repository's existing unittest CI does not
automatically collect it.

Related: #9510 refactors the training loop but retains the configured-size
loss division in these implementations; this fix is independent of that work.