Train: add auto-DDP distributed training

[GeorgePearse]

Summary

• Port the automatic single-node DDP spawning approach from BinItAI/core PR #5414.
• Fill in the missing visdet.engine.dist implementation (rank/world_size, init, collectives, result collection), which is required by Runner/samplers/metrics.
• Add visdet.engine.runner.auto_train() and a small entrypoint script scripts/train_auto_ddp.py (no torchrun required).

Usage

• python scripts/train_auto_ddp.py path/to/runner_config.py

Testing

• python3 -m pytest -q tests/test_runtime/test_dist_env_fallback.py

[github-actions]

🤖 Multi-Model Consensus Review

Model	Status
Claude Sonnet 4	✅
GPT-4o	✅
Gemini 2.0 Flash	✅

✅ scripts/train_auto_ddp.py

Pass Rate: 3/3 models

✅ tests/test_runtime/test_dist_env_fallback.py

Pass Rate: 3/3 models

✅ visdet/engine/dist/__init__.py

Pass Rate: 3/3 models

✅ visdet/engine/dist/dist.py

Pass Rate: 3/3 models

✅ visdet/engine/dist/dist_utils.py

Pass Rate: 3/3 models

✅ visdet/engine/dist/utils.py

Pass Rate: 3/3 models

✅ visdet/engine/runner/__init__.py

Pass Rate: 3/3 models

✅ visdet/engine/runner/auto_train.py

Pass Rate: 3/3 models

[github-actions]

✅ Skylos Scan: No dead code or security issues detected.