Add distributed training support for multi-GPU/multi-node synthetic data generation

[Oludiran-Ayoade]

Feature Description

Add support for distributed training and generation to handle datasets that exceed single-GPU/node memory limits.

Motivation

Currently, Synthcity plugins are limited to single-GPU training, which restricts the size of datasets that can be processed. Large healthcare, financial, or scientific datasets often require distributed computing resources.

Proposed Solution

Implement a distributed training wrapper that supports:

• Multi-GPU training using PyTorch DDP (DistributedDataParallel)
• Distributed data loading with automatic sharding
• Gradient accumulation for memory-constrained environments
• Distributed generation with load balancing
• Integration with Ray or Dask for cluster orchestration