ScheduleFree AdamW Support

[ertugrul-dmr]

This pull request adds support for a new "schedule-free" optimizer as described in the paper Schedule-Free Optimizer for Fast Convergence. This optimizer can be easily integrated into any training pipeline without breaking existing functionality.
Usage

To use this optimizer, simply add the following configuration to your training setup:

lr: 5e-4
optimizer: "schedulefree"
optimizer_params:
    weight_decay: 0.1

While this optimizer promises faster convergence as per the paper, preliminary results suggest it may overfit or "burn out" quickly in certain cases. Optimal parameters are yet to be determined, so further experimentation may be required to fine-tune for specific tasks.