feat: Add checkpoint forking functionality

[corbt]

Summary

This PR adds the ability to fork checkpoints from existing models, which is useful when training goes off the rails and you want to restart from a previous checkpoint with different parameters.

Changes

• Added _experimental_fork_checkpoint method to Backend and LocalBackend classes
• Changed S3 pulling API from latest_only boolean to only_step parameter that accepts either an int or "latest"
• Added debugging output to help diagnose S3 sync issues

Example Usage

# In project_types.py, configure fork settings
class ProjectPolicyConfig(BaseModel):
    fork_from_model: str | None = None
    fork_from_project: str | None = None  
    fork_not_after_step: int | None = None

# In all_experiments.py, create a forked model
models["229"] = models["224"].model_copy(deep=True)
models["229"].name = "email-agent-229"
models["229"].config.fork_from_model = "email-agent-224"
models["229"].config.fork_not_after_step = 1381

# In train.py, fork will happen automatically if configured
if model.config.fork_from_model:
    await backend._experimental_fork_checkpoint(
        model,
        from_model=model.config.fork_from_model,
        from_s3_bucket=os.environ["BACKUP_BUCKET"],
        not_after_step=model.config.fork_not_after_step,
        verbose=True,
    )

Key API Changes

only_step parameter on

• only_step=None - Pull all checkpoints (default)
• only_step="latest" - Pull only the latest checkpoint
• only_step=1234 - Pull only checkpoint 1234

[corbt]

The good news is that forking seems to work. (The bad news is that my forked run failed in almost the same place, but you can't win 'em all).

[corbt]

@bradhilton I think this is ready for review!