Continue DDP implementation from existing branch ddp-support-wip

[guillaume-byte]

Summary

Continue and complete the existing Distributed Data Parallel (DDP) implementation work from branch ddp-support-wip.

Scope

• Review and continue current implementation from ddp-support-wip.
• Validate end-to-end DDP training flow in WeightsLab.
• Ensure behavior parity with single-GPU baseline for core metrics/logging/checkpointing.
• Document launch/config usage and known limitations.

Acceptance Criteria

• DDP path is runnable from documented command/config.
• Multi-GPU training completes for at least one reference workload.
• Metrics, logging, and checkpointing are validated under DDP.
• Existing branch work is merged (or superseded) with migration notes if needed.

Branch Reference

• Existing work branch: ddp-support-wip