Feature Request: Add FSDP (Fully Sharded Data Parallel) Support

Great work on the repository! Thanks for already supporting DDP.

To better support training of large models, could you please consider adding support for FSDP?

FSDP shards model parameters across GPUs, which can dramatically reduce memory usage compared to DDP. This is especially helpful for users with limited VRAM who want to train larger models.

Thanks for considering!