fix: lazy-import heavy deps in CLI so lightweight commands work without extras

[Kovbo]

Summary

• CLI commands (art install-skills, art train-sft, art train-rl, art migrate) crash with ModuleNotFoundError when only base dependencies are installed, because cli.py eagerly imports fastapi, uvicorn, torch (via LocalBackend), and numpy (via model.py → calculate_step_metrics) at module load time — even though only the run command needs them.

• Moved all run-only imports (fastapi, uvicorn, pydantic, LocalBackend, dev, ARTError, Model, TrainableModel, TrajectoryGroup, TrainConfig) from top-level into the run() function body.

• Made the calculate_step_std_dev import in model.py lazy (moved into the method that calls it), breaking the __init__ → utils → output_dirs → model → numpy eager import chain.

• Removed unused imports: Optional, Provider, TogetherDeploymentConfig, WandbDeploymentConfig.

Reproduction

pip install openpipe-art
art install-skills
# ModuleNotFoundError: No module named 'numpy'
# (after installing numpy) ModuleNotFoundError: No module named 'fastapi'
# (after installing fastapi) ModuleNotFoundError: No module named 'torch'

Test plan

• pip install openpipe-art && art install-skills works with base deps only
• art run still works when backend extras are installed
• ruff check passes (verified locally)

🤖 Generated with Claude Code