Forward SFT gradient accumulation to MLX-LM

[kargarisaac]

Summary

• forward SFTConfig.gradient_accumulation_steps into MLX-LM TrainingArgs
• print the effective native-training gradient accumulation value in the SFT config log
• add a mock-based regression test that verifies native SFT training passes the configured value through

Why

SFTTrainer already reads and stores gradient_accumulation_steps, but the native MLX training path did not pass it into mlx_lm.tuner.trainer.TrainingArgs, so MLX-LM used its default of 1.

This makes the Unsloth-compatible SFTConfig(gradient_accumulation_steps=...) behave as expected for native SFT training.

Tests

.venv/bin/python -m pytest tests/test_trainers.py -q
# 16 passed

.venv/bin/python -m py_compile mlx_tune/sft_trainer.py tests/test_trainers.py
git diff --check

Notes:

• uv run pytest tests/test_trainers.py -q currently fails during dependency resolution before tests run because pyproject.toml allows Python >=3.9, while mlx>=0.31.0 requires Python >=3.10.
• ruff check mlx_tune/sft_trainer.py tests/test_trainers.py reports pre-existing style issues in these files, so I did not broaden this bugfix PR into a formatting cleanup.