feat: add persistent state methods to Model

[corbt]

Summary

Add write_state() and read_state() methods to art.Model that persist arbitrary JSON state to the model's output directory (state.json).

Features

• StateType type parameter with default dict[str, Any] for backward compatibility
• write_state(state: StateType) - persists state as JSON to {output_dir}/state.json
• read_state() -> StateType | None - reads state, returns None if not found
• Full backward compatibility: existing Model[MyConfig] syntax still works
• Optional type safety: Model[MyConfig, MyState] enforces state type with type checkers

Use Case

This enables filesystem-based state tracking for training resumption, dataset position, and other metadata. For the Tinker RL training migration, this replaces wandb-based state tracking with filesystem-based persistence for better compatibility.

Usage Examples

# Without type parameter (backward compatible)
model = art.TrainableModel(name="my-model", project="my-project", base_model="...")
model.write_state({"step": 5, "dataset_offset": 100})
state = model.read_state()  # Returns dict[str, Any] | None

# With type parameter for type safety
class TrainingState(TypedDict):
    step: int
    dataset_offset: int
    policy_version: int

model = art.TrainableModel[TinkerConfig, TrainingState](...)
model.write_state({"step": 5, "dataset_offset": 100, "policy_version": 3})
state = model.read_state()  # Returns TrainingState | None

File Storage

State is stored at: .art/{project}/models/{model_name}/state.json

[bradhilton]

LGTM