0.3.5

• More mlx synchronization: grad checkpoints, updates to batching functionality
• wandb_sweep.py -f/--prompt-format and --train-type arguments: Support for completion-only sweeps
• Fix reference to internal temperature argument
• Add 'min_cos_lr' to Cosine with warmup schedule.
• Add update of progress bar in training callback
• Revert to using iterate_batches from mlx_lm. Fix use of progress bar and new datasets types with self-supervized training.

Full Changelog: https://github.com/chimezie/mlx-tuning-fork/commits/0.3.5